The crew schedule is a crucial component of urban rail transit operation plannings， and its optimization methods have become a hot research topic in the field of transportation. To promote the research progress in crew planning optimization for urban rail transit， reviewing relevant literatures from both domestic and international sources. Regarding mathematical models， systematically summarizing two typical modeling approaches which are respectively based on "set relations" and "network graph construction"； Concerning solution algorithms， provideing detailed assessments of traditional integer programming algorithm， mathematical programming method based on column generation， graph and network-related algorithm， and intelligent heuristic algorithm. Finally， addressing current research deficiencies and challenges to propose future research directions.

The current research status in the field of takeover behavior was summarized from two aspects： impact mechanism and improvement methods. In terms of the influence mechanism of takeover behavior， the influence factors were systematically divided into automated driving system factors， traffic factors and driver factors， and more carefully summarizes the influence mechanism of driver factors. With respect to the methods for improving takeover behavior， based on the conclusions on the impact mechanism of takeover behavior， a series of methods were summarized from aspects such as optimization design for human-machine interaction， takeover behavior modeling and prediction， and driver training for takeover. Finally， the current problems and future research directions were proposed from the perspectives of influence mechanisms and improvement methods.

The catalytic burner and heat exchanger in the peripheral thermal management system of a compact solid oxide fuel cell are coupled into a catalytic heat exchanger （CHE）. The thermal performance of the CHE and the catalytic conversion of methane were investigated experimentally. The performance of the CHE was tested， compared and analysed when only catalytic combustion or heat exchange was taking place， compared to the performance during normal operation. The results show that the heat transfer efficiency of the CHE in normal operation is 10% higher than when only heat exchange is carried out， while the catalytic conversion of methane is 4% lower than when only catalytic reaction is carried out. Although CHE reduces some of the catalytic conversion performance， it improves fuel utilisation， so using CHE to replace the burner and heat exchanger in a system can improve heat transfer performance， speed up system start-up or be more fuel efficient for the same start-up time， save manufacturing costs and system space， and reduce the heat load on the remaining thermal components. Potential applicability and broad prospects in compact SOFC systems.

A parameter design method of multiple parallel dynamic vibration absorbers is proposed in order to suppress the powertrain system torsional resonance that exists in various gears of automotive. First， a four degrees of freedom powertrain model is presented， and the expression of the frequency response functions of the vibration system are derived based on the equivalent system with n parallel dynamic vibration absorbers attached. Then， the Nondominated Sorting Genetic Algorithm II is adopted to optimize the frequency ratios and damping ratios of the four parallel dynamic vibration absorbers in order to minimize the angular displacement and acceleration of the vibration system under three different gear ratio conditions. The Technique for Order Preference by Similarity to an Ideal Solution combined with the Entropy Weight Method is utilized to sort the Pareto solutions. Finally， the proposed optimization method is validated effective by comparing with three other traditional methods， and the time and frequency domain simulations are also implemented for the validation. The proposed method can provide references for the optimal parameter design of the dynamic vibration absorber when the eigenvalue of the target system is variable.

Continuous test scenario is an important content of automated driving vehicle scenario-based testing system， but its fairness has been disputed. For this reason， this paper proposes a real-time evaluation method for the continuous test scenario complexity of automated driving vehicle， with a view to solving the problem of test scenario fairness. Based on the six-layer scenario model， a scenario element importance assessment system is established； through analyzing the mapping relationship between scenario elements and perception， decision-making， and execution systems， a quantitative evaluation method of scenario complexity at the system layer is constructed； the impact transfer weighting coefficients of different system are calculated， which can be used to realize the comprehensive calculation of scenario complexity in real time. The traffic circle continuous test scenario is selected to analyze the scenario complexity of two black-box automated driving systems， the results of the scenario complexity of the two vehicles are 1 and 0.765， which are consistent with the scenario complexity encountered by the two tested systems in the test process. The results can prove the effectiveness of the method proposed in this paper.

For the 1.5 mm thick AA6061-T6 aluminum alloy sheet for bodywork， the effect of two processes on the tensile shear performance of PFSSW-glued spot welding joints was studied using the penetrating glue welding method and the capillary action glue welding method. The temperature field in the lap region was analyzed by the thermal coupling method， and then the tensile shear properties and fracture modes of the joints using different process sequences were compared. The results show that the capillary action glue welding method can significantly improve the tensile shear performance of aluminum sheet joints compared to PFSSW and penetration glue welding methods. Methodological support is provided for the joining process of aluminum sheet metal and other glue welded joints for bodywork.

Aiming at the problem that the design of the non-pulsating stator curve is limited when the argument distribution coefficient of the inner curve radial piston motor is small， a method for non-pulsating argument distribution according to the principle of argument superposition is proposed. Under the low distribution coefficient， the non-pulsation argument distribution of seven kinds of motion law curves， such as equal acceleration， equal acceleration with transition region， sinusoid， sinusoidal with transition region， parabola， parabola with transition region and uniform acceleration correction equal acceleration were studied. Taking the radial piston motor with six-acting eight-piston inner curve as an example， the influence of distribution result and argument on output characteristics is analyzed by Matlab/Simulink. The results show that ： ①This method can realize the non-pulsation angle distribution of each curve with transition zone under low distribution coefficient， and the zero velocity zone should be as small as possible when the working condition is satisfied. ②For the uniform acceleration correction equal acceleration curve， the front and rear correction angles should also be small， and the front correction angle can be larger than the rear correction angle to make the contact stress change more uniform.

For the fault diagnosis on rolling bearing， a method via wavelet packet logarithmic-energy map is proposed. Firstly， a new wavelet packet node logarithmic energy formula is improved and presented to overcome the complexity and subjectivity of parameters in the traditional ones. Thus the high-frequency faults and the low-frequency faults are easily identified. As a result， the initial time-frequency features are extracted adequately. Secondly， the idea of Gramian angular summation field is used to transform the features from one-dimension data to two-dimension picture. Therefore the features via wavelet packet logarithmic-energy map are constructed. In them， the space information among the adjacent features are considered further. So the optimization for the initial time-frequency features is completed and their significance are increased. On this basis， the residual network is applied to enhance the accuracy of the presented approach. Finally， the higher accuracy of diagnosis and the greater generalization ability of the proposed method is verified by the standard rolling bearing data set of Case Western Reserve University.

With the development of the hydrogen fuel cell vehicles industry， the control of proton exchange membrane fuel cell engine has become the focus of research. To improve the life and efficiency of fuel cell engines， the air supply pressure should be precisely controlled. Therefore， we primarily research the control of air supply pressure in this paper. Firstly， the modeling of air supply pressure is established. Secondly， considering the existence of model uncertainty and disturbances， we design the disturbance observer to estimate and compensate this influences. Thirdly， a new slide mode control method is proposed to achieve accurate pressure control based on the established mathematical model. Finally， the effectiveness of the air supply pressure control strategy is verified through four simulation conditions.

To evaluate the effectiveness of short video dissemination on traffic congestion alleviation， this paper takes Tik Tok short video dissemination as the research object， and through the analysis of the dissemination characteristics of short videos， selects 5 major elements and 13 dimensions to construct an indicator system for short video characteristics. Based on above， a random forest model of machine learning is constructed to predict effective probabilistic of traffic short video dissemination and compare it with traditional prediction models. In particular， this paper proposes the concept of traffic diversion validity index for the first time， and quantitatively analyzes the effectiveness of short video on traffic diversion. Taking the road in the Five Avenue of Tianjin as an example， the results show that the six dimensions of the source， theme， material， purpose， communication power and comment content guide of the short traffic video have a strong correlation with their effective communication probability， and the average effective communication probability is in the range of 0.19-0.28； the final calculation results of the short video traffic guidance effectiveness index of Line 1 and Line 2 are 0.008 pcu/person-time and 0.011 pcu/person-time， respectively. Generally speaking， it is considered that the effect of traffic congestion relief is obvious. This paper provides new ideas for road traffic control in the new media era.

To investigate the heterogeneity and transferability of factors influencing accident severity， a random parametric Logit model based on mean and variance heterogeneity is constructed to analyze the factors influencing the severity of injury in lane-changing accidents. Divide the severity of accidents into three categories and select 27 potential influencing factors from four aspects： driver characteristics， vehicle characteristics， road characteristics， and road environmental characteristics. Capturing heterogeneity in factors affecting accident severity through mean and variance changes in stochastic parameters，testing the transferability of factors influencing accident severity using log-likelihood ratios，quantifying the impact of factors on accident severity through marginal effects. The results show that there is significant heterogeneity in the factors influencing the severity of accidents in the two groups of lane change and non-lane change. Under the condition that there is no correlation between the various influencing factors， sunlight exposure is a random factor in lane changing accidents， its mean is significantly correlated with the driver's speed change and interstate highways， and its variance is significantly correlated with the complete failure of vehicle disabling functions. Signal control is a random factor in non lane changing accidents， and its mean is correlated with the driver's age as middle-aged， and its variance is correlated with the driver's speed change. There is no transferability and significant difference in the factors influencing the severity of lane change and non-lane change accidents.

To meet the need of standard open datasets for the researches of roadside multi-source fusion sensing algorithms， this paper proposed a method to construct a standard roadside multi-source sensing dataset. The LIDAR and image data was collected at an urban T-junction and matched each other in both spatial and temporal dimensions. A vehicle three-dimension configuration extraction method was proposed， which included the steps of road space division， road pavement segmentation， and laser point cloud clustering. A vehicle labeling method was designed， which included the steps of target filtering and classification， recognition difficulty division， 3D bounding box calibration， and tag information supplement. It constructed a standardized roadside multi-source sensing dataset that contained the labels of 9 794 cars and heavy vehicles in daylight and nighttime. The YOLOv5 algorithm and the PointRCNN algorithm were used to test the 2D and 3D target recognition performance on the constructed dataset. Test results showed that due to the differences of scene complexity， data collection device and vehicle type， the constructed dataset and open vehicular datasets had differences in terms of average number of scene and vehicle laser points， and size of vehicle 3D bounding box. The YOLOv5 algorithm and the PointRCNN algorithm have similar vehicle target recognition accuracy on the open vehicular datasets and the constructed roadside multi-source sensing dataset.

The uneven spatial-temporal distribution of passenger demand affects the level of public transportation service. The emerging modular vehicle system can adapt to the spatial-temporal demand changes by changing the number of modular vehicle units flexibly. Therefore， a zonal-based flexible feeder bus routes optimization model considering the modular vehicle system is established. The model aims at minimizing the total system cost consisting of vehicle operation cost， fixed cost and passenger travel time cost， and solves the model using a multi-agent genetic algorithm combining genetic algorithm and multi-intelligence system. Finally， numerical experiments based on the service area and stops of Xi'an “Jie Bus” are designed. The results show that the total cost can be reduced by about 18.31% by considering the modular vehicle system compared with the traditional fixed-capacity feeder bus system， which provides a new idea for the future development of urban public transportation.

In order to explore the impact of built environment on residents' commuting mode choice under the mixed distribution of different planning unit groups， a mixed logit model considering the heterogeneity of housing type spatial units was constructed. At the same time， the impact of built environment characteristics at both individual and planning spatial unit levels on commuting behavior was considered. An empirical study was conducted using sample data from a survey of residents' transportation behavior and willingness in Shenzhen. The results show that the mixed logit model has better goodness of fit than the multinomial logit model. If the differences in housing spatial unit groups are ignored， the degree of influence of land use mix on mixed housing spatial units will be underestimated by 18.5%. The degree of influence of rail transit stations on affordable housing， small property rights housing and resettlement indicator housing spatial units will be overestimated by 6.3%， 5.4% and 4.5%， respectively， while the influence of bus stop density on resettlement indicator housing spatial units will be underestimated by 4.8%. Public service facilities have the greatest empirical sensitivity to commercial housing （3.2%） and small property rights housing spatial units （2.8%）， while bus stop density has the greatest empirical sensitivity to affordable housing （2.6%） and resettlement indicator housing （2.0%）， and distance to the nearest rail transit station has the greatest empirical sensitivity to small property rights housing spatial units （2.1%）. Therefore， it is necessary to combine the impact and sensitivity differences of different housing spatial units' built environment， and accurately guide residents' green commuting through improving high-efficiency housing spatial units' built environment.

In order to improve the mixed traffic of connected automated vehicles and human-driving vehicles， this paper proposed a collaborative optimization method for signals and trajectories of connected automated vehicles on dedicated bus lanes. In the double-layer optimization model of signal and trajectory， this paper builds linear mathematical relations between trajectory variables and traffic signal variables， and compute the solutions in sections of time and space. Consequently， this model is used in real number case to verify the validity and operate comparative experiments between collaborative and non-collaborative optimization. It proves that efficiency of connected automated vehicles in collaborative optimization is 7.7% higher.

To explore the complete commuting mode choice behavior when residents are aware of PM_2.5 inhalation and risk under different polluted weather， a revised ordering-based max-relevance and min-redundancy greedy （OMRMRG） Bayesian Network model was established to analyze the chain-effect utility of various variables. A questionnaire was designed to obtain residents' travel intentions in polluted weather after informing them of the risk level of their current commuting patterns and the pollutant inhalation of different travel modes. An OMRMRG structural learning algorithm with the introduction of mutual information theory was proposed to explore the chain-effect utility based on the causal relationship between variables. The results of analyzing the survey data of Xi'an residents show that： the variable chain with the greatest influence under light pollution is age-the presence of private car-daily commuting mode-polluted weather commuting mode， and 39% of residents reselect a commuting mode with a lower risk level than the daily commuting mode. The chain-effect utility of the influencing factors under heavy pollution is shown as age-income-the presence of private car-polluted weather commuting mode， and the correlations between the adjacent variables are positive， positive， positive and negative， respectively. The study reveals the chain-effect utility of different pollution conditions affecting residents' commuting patterns， which is beneficial to develop more constructive induced strategies to help residents travel healthily.

To analyze the driving risk of vehicle braking behavior on curved sections of superhighways， the impact of driving speed， radius of circular curves， road superelevation， and braking force on driving risk is comprehensively considered. The vehicle dynamics model is established， and the lateral acceleration and yaw rate are taken as the evaluation indicators of driving risk. The influence of various factors on driving risk is analyzed. The research results indicate that the radius of circular curves has a significant impact on the driving risk of curve sections on superhighways， with the primary and secondary order of influencing factors being circular curve radius， speed， braking force， and road superelevation. When conducting risk assessment on vehicles driving on curved sections of highways， it is necessary to comprehensively consider the lateral acceleration and yaw rate of the vehicle. The research results can provide reference for analyzing the safety of vehicles driving on superhighways.

In response to the lack of dynamic characterization of residents as the main body of urban spatial activities in traditional functional area identification methods， this paper proposes a functional attribute identification method for urban parcels based on taxi trajectory and POI data. Firstly， a class of departure and arrival time vectors are constructed respectively. Then， the residents' travel patterns is clustered by an improved dynamic time regularization and clustering algorithm. Finally， the functional attributes of the blocks are identified by combining the residents' travel curve characteristics， POI density and enrichment index. Taking Xi'an city as an example， the departure and arrival pattern characteristics of residents in different regions on weekdays and rest days are discussed to identify the functional attributes of different blocks within the city. The results show that different departure and arrival patterns represent different peaks in the morning peak， afternoon peak， evening peak， night and early morning， and the corresponding blocks show a certain circle structure in the spatial distribution and their respective functional tendencies. The identification of the functional attributes of the parcels using the departure-arrival pattern characteristics of residents and POI information has a complementary effect， and the functional attributes show a ternary structure of "employment-residence-rest"， which also reflects the spatial and temporal changes of different functional areas and people's activities. The results of the study are useful for planning departments to reallocate transport resources and optimize the spatial structure of the city.

To avoid frequent start-stop of vehicles and reduce delays and fuel consumption at exit lane for left-turn intersections， an optimal trajectory control model for left-turning vehicles is proposed. The model is applicable to exit lane for left-turn intersections in an intelligent connected vehicle environment. Firstly， the unique features of the exit lane for left-turn intersections are analyzed in terms of fuel consumption compared to traditional intersections. Secondly， the corresponding trajectory control double-layer model is established by considering the geometric design and traffic signals of the intersection， optimizing the lane selection and the time to enter the exit lane for left-turn. Finally， SUMO and Python software are used to simulate and verify the proposed control model and simulated experiments are designed to analyze the effects of different parameters such as traffic volume， green signal ratio and length of exit lane for left-turn on the model's performance. The experiments show that the proposed optimal trajectory control model effectively reduces delays and fuel consumption， with an average reduction of 26.6% in vehicle delay and 50.2% in fuel consumption compared to the Krauss control model. Moreover， the optimal trajectory control model exhibits robustness under different lengths of the exit lane for left-turn and has wide applicability.

To achieve accurate identification of the driving intentions of surrounding vehicles by intelligent connected vehicles， a driving intention recognition framework based on trajectory prediction and extreme gradient boosting （XGBoost） algorithm is proposed. Firstly， we attach the driving intention label to the vehicle's historical trajectory sequence and build the offline training dataset. Then， a driving intention recognition framework is constructed. A mixed teacher force long short-term Memory （LSTM） module is used to predict the future trajectory. The XGBoost module concatenates the historical and future trajectory and recognizes driving intention （left lane change， lane keeping， and right lane change）. Finally， the model is verified on the real road datasets Next Generation SIMulation（NGSIM） US101 and I-80 sections. The experimental results show that the proposed model outperforms the other methods in metrics precision， recall， F1 score， and accuracy. The recognition accuracy can reach 97.7% in the prediction of 4 s historical trajectory and 3 s future trajectory， which shows good performance in driving intention recognition. The code can be obtained at：https：∥gitee.com/fanghz-colin/lstm-xgboost.git.

In order to effectively solve the problems of poor vehicle scheduling scheme and low satisfaction of materials in multi-distribution center， a multi-objective ant colony algorithm based vehicle scheduling optimization method for multi-distribution center emergency materials distribution was proposed. Considering the actual situation of the distribution route， aiming at maximizing the average transportation time and material satisfaction， the multi-objective function of vehicle scheduling optimization for emergency materials distribution in multi-distribution centers was constructed. Based on the multi-objective ant colony algorithm， the optimal vehicle scheduling scheme for emergency materials distribution in multi-distribution center is determined. The experimental analysis proves that the proposed multi-distribution center emergency materials distribution scheme is more reasonable， the scheduling comprehensive coefficient is above 0.950， and the material satisfaction rate is above 90.00%， which effectively improves the vehicle scheduling effect and can develop a more reasonable scheduling scheme.

In view of the limitations of using macroscopic homogeneous medium method and laboratory test to study the performance of asphalt mix， a numerical modeling method for meso-scale of asphalt mix is proposed. This study uses discrete element PFC software to establish aggregate with real morphology， and the discrete element models of asphalt mix were established by the scaling method. The meso-parameters of the Burgers model between asphalt mortar particles were determined by using the dynamic shear rheology （DSR） test and the bending beam rheology （BBR） test combined based on the time-temperature equivalence principle. The distribution of tensile and compressive stress in the specimen during the uniaxial compression test and Marshall splitting test. The results show that irregular morphology of Clump aggregate can be generated by filling the bounding box with ODEC algorithm. The radius of the floating points obtained by traversing the specimen is enlarged so that the distance between the floating points and the adjacent particles is reduced to generate contact. This method can be used to eliminate the floating points inside the discrete element specimens of asphalt mixes. The asphalt mix model built by using the discrete element software PFC can well simulate the mechanical behavior of the asphalt mix at low temperature. The results have great significance to carry out meso-mechanics research and numerical modeling calculation of asphalt mix.

To explore the effect of mesoscale structure of aggregates in asphalt mixtures on the macroscopic properties， the secondary shape parameters length-width ratio and width-thickness ratio were used to characterise the real aggregates was proposed in this paper using X-ray CT. On the basis of the secondary shape parameters， the aggregate was reconstructed used an ellipsoidal surface base， the specimens was generated for splitting tests in PFC3D using the servo control. The results show that the angularity aggregates model had the highest accuracy when the length-width ratio was taken as 1.5 and the width-thickness ratio was taken as 0.98. In the reconstruction of angularity aggregates， the length-width ratio should avoid exceeding 2.0. Elongate aggregates hindered the formation of a stable aggregate skeleton， flat aggregates made the tension in the centre of the mixture not obvious. Well angularity aggregates improved the tensile properties. The ellipsoidal surface based of the aggregate reconstruction method was an accurate and efficient modelling method.

Analytical research on the horizontal thrust influence line of the variable cross section two-hinged arch is still imperfect， this study focuses on the derivation and application of the analytical solution of the thrust influence line of the variable cross section two-hinged arch under parabola and catenary. Ritter's formula is used to establish the variation rule of the arch ring section， and the curve integration of the variable cross section two-hinged arch is carried out. Put forward on the basis of the principle of force method of two kinds of linear variable cross-section two hinged arch horizontal force influence line analytical solution， compared with the finite element results show that the proposed formula and finite element calculation error within 8.5%， and then extract the two foot arch structure before and after damage force influence line difference curvature damage identification index， based on force influence line， a new method of variable cross-section two hinged arch damage identification， The finite element example verifies that the proposed method can realize the damage location of the two-hinged arch structure， which can be used as the basis for the arch foundation strength design under the action of moving loads and the theoretical reference for the rapid detection application of arch Bridges.

The process of soil frost heave is nonlinear， and the stress deformation of the lining structure is also nonlinear. Under the action of frost heave force， the structure of the channel will undergo nonlinear deformation， including bending， shear， and torsion， which is particularly prominent in channels in cold regions. Therefore， it is necessary to accurately simulate and analyze the process of frost heave. In this context， a study is conducted on the frost heave mechanical model of lining channels in cold regions under the theory of double layered beams on elastic foundations. This study sets basic assumptions in order to simplify and abstract the problem in a reasonable manner， in order to better establish mathematical models. Using the theory of double layered beams on elastic foundations， the lining channel can be simplified into an elastic double layered beam structure and three mechanical equations， namely deflection equation， bending moment equation， and shear force equation， can be established to describe the deformation of the double layered beam structure. Under boundary conditions， use finite element analysis to solve equations and complete the mechanical analysis of frost heave in lining channels in cold regions based on the solution results. The experimental results show that the frost heave damage of U-shaped lined channels mostly occurs on both sides， and as the frost heave amount increases， the deflection and bending moment of the lined channel gradually increase， while the shear force gradually decreases. However， when the frost heave amount increases to a certain extent， the deflection， bending moment， and shear force sharply decrease.

Aiming at the problem that the existing neural network methods are not accurate enough to predict the Killip classification of patients with myocardial infarction， a weighted isomorphic graph classification algorithm that can learn causal features is proposed. Use the differentiated learning objectives to separate the graph-level representation features of causal correlation and non-causal correlation， and then use the backdoor adjustment method in causal reasoning to reduce the confusion of non-causal features on classification results. The experimental results show that the average accuracy of the method proposed in this paper reaches 80.52% in the prediction task of the Killip grading of patients with myocardial infarction， which is 3.72% higher than that of the non-graph neural network method， and 0.72% higher than that of the graph neural network method without learning causal characteristics. Therefore， the method proposed can better complete the task of predicting the Killip grading of patients with myocardial infarction.

To ensure the security and privacy of fingerprint information， a very few sample conditional fingerprint image detail encryption algorithm considering privacy information security is proposed. Under extremely few sample conditions， the fingerprint plaintext image is divided into multiple square image blocks using the block principle of the square theorem to construct the magic square matrix required for scrambling. After unfolding and scrambling each matrix， operations such as stitching and transposing are carried out to fully diffuse all pixels and restore them to the original fingerprint image size. After multiple scrambling operations， the pre encrypted fingerprint image is obtained. Compose a fingerprint image cube from plaintext images， form an initial key through the hash value and external parameters of the plaintext image， and substitute it into a chaotic system to form the desired sequence. Using chaotic sequences to form masks and rule matrices， dynamic DNA encoding is performed on the mask matrix and scrambling results. At the same time， DNA operations are performed on the encoded image cube and mask matrix through a rule matrix， and the final encrypted fingerprint image details can be obtained through DNA dynamic decoding. The experimental results show that the proposed algorithm can more efficiently ensure the privacy information security of fingerprint images under very few sample conditions.

In order to improve the effectiveness of detecting illegal vehicles in traffic monitoring videos， a multi-objective detection method for illegal vehicles in traffic monitoring videos based on the YOLOv5 network algorithm is proposed. In the fusion processing of traffic monitoring video images， the grayscale histogram equalization operation is performed， and the grayscale deviation values of different exposed images are calculated through the window function. All ghosts in the fused images are removed by image corrosion， and high-quality image fusion results are obtained through pixel normalization and Laplace pyramid； Based on the YOLOv5 network algorithm， a multi head self attention learning mechanism is constructed using a driver graph encoder in the Transformer module to enhance the semantic information of target features of illegal vehicles in the image； On the basis of continuous dilated convolution， dense connection structures are utilized to expand the feature map convolution into a single pixel addition operation， enhancing the semantic information of features； Utilizing the Softmax function for feature multi-scale fusion to achieve multi-target detection of illegal vehicles in traffic surveillance videos. The experimental results show that the proposed method can effectively improve the quality of traffic monitoring video images， and the YOLOv5 network algorithm used has high computational intensity， resulting in more accurate multi-target detection of illegal vehicles.

The social recommendation algorithm based on graph neural networks has achieved good performance in improving the performance of recommendation systems. However， existing methods overlook the dynamic evolution of users interest preferences and items attractiveness， as well as the potential connections between items. This can lead to the model learning features that are not accurate and rich enough， limiting prediction accuracy. For this problem， a social recommendation model based on the global capture of dynamic， static and relational features is proposed. The model first captures the long-term static features and short-term dynamic features of users and items through interaction modeling network and temporal modeling network， respectively. Then， the gated fusion network adaptively fuses the long-term static features and short-term dynamic features to obtain the dynamic and static features. Finally， the relationship aggregation network is used to capture the relational features. The experimental results on the Ciao and Epinions datasets show that the prediction error of the proposed model is significantly lower than that of the existing advanced methods， and it has good application value.

To address the issues of target omission and the inclusion of a large number of background points in keypoint sampling for point cloud-based object detection， an improved algorithm based on the PV-RCNN network is introduced. This approach employs both a regional proposal fusion network and weighted non-maximum suppression （NMS） to merge proposals generated at various scales while eliminating redundancy. A segmentation network is utilized to segment foreground points from the original point cloud， and object center points are identified based on these proposals. Gaussian density functions are employed for regional density estimation， which assigns different sampling weights to solve the problem of difficult sampling in sparse areas. Experimental evaluations on the KITTI dataset indicate that the algorithm enhances the average precision at medium difficulty levels by 0.39%， 1.31%， and 0.63% for cars， pedestrians， and cyclists， respectively. Generalization experiments were also conducted on the Waymo open dataset. The results suggest that the introduced algorithm achieves higher accuracy compared to most of the existing 3D object detection networks.

Considering the small size of calcified plaques and the difficulty in distinguishing them from non-plaque areas， a medical prior knowledge-guided optimization method for proposal generation was proposed. This method is based on the object detection network Faster R-CNN. Proposal generation is optimized through the location and shape of anchor. Calcified plaque definition is used to generate a guidance mask image to direct the location generation of proposal. Anchor shape prediction branch is employed to optimize the shape generation of proposal. A multi-scale guidance mask pyramid architecture is proposed for feature maps of different scales in FPN. The experimental results of detecting calcified plaques in CCTA images show that the proposed method improved AP and Recall by 12.8% and 25.7% respectively compared with Faster R-CNN with standard RPN； When the average number of false positives per image is 2， Recall of the proposed method is 86.05%.

To address the issues of lacking the modeling of long-range multidimensional dependencies of skeleton joints and weak temporal feature extraction capability in existing skeleton-based action recognition methods， resulting in low recognition accuracy and poor generalization capability， a hyper-connected graph convolutional network （HC-GCN） action recognition model was proposed. First， the basic working principle of the adaptive graph convolutional network was introduced； second， the hyper-connected adjacency matrix construction method was proposed and combined with the multidimensional adaptive graph convolutional network module， and then the hyper-connected adaptive graph convolutional network （HC-AGCN） module was constructed. Then， the residual connections were introduced in the omni-dimensional dynamic convolution to create the residual omni-dimensional dynamic temporal convolutional network （ROD-TCN） module， and the HC-GCN model was further proposed by combining the HC-AGCN module and trained under a two-stream-three-graph network. Finally， the validation experiments on the performance of the model based on the NTU-RGB+D and NTU-RGB+D 120 datasets were conducted. Experiment results reveal that the recognition accuracy of the model on the datasets stated above is 96.7% and 89.0%， respectively， demonstrating the model's exceptional accuracy and generalizability.

To timely obtain video quality， a lightweight detection algorithm for lossless transcoding and heavy compression of massive digital media videos is proposed. This method divides video encoded frames into several encoding tree blocks based on high-performance video encoding， and optimizes video rate distortion to ensure video quality during the partitioning process while minimizing computational and storage costs， achieving a certain degree of lightweighting； Based on this， pixel values， color space， and motion vectors are extracted from each image frame in the video， and the extracted results are compared and analyzed with the initial video pixel values， color features， and motion vectors to determine whether the quality of the re compressed video is damaged， achieving lightweight detection of the re compressed video. The experimental results show that when using the proposed method for compressed video quality detection， the color component extraction results of the compressed video image are consistent with the actual color component of the compressed video image. Moreover， when the number of videos reaches 5000， the detection result of the number of false positives in video detection is 3， further demonstrating the high detection performance and good effect of the proposed method.