With the increasing system complexity in various engineering applications， greater demands are being placed on the reliability and safety of these systems. A kind of more comprehensive requirement called “health” has been naturally put up beyond reliability and safety. Consequently， Prognostics and Health Management （PHM） for complex systems has naturally become a hotspot in systems engineering. Moreover， it has been applied in many areas， such as aerospace， machinery， and power electronics. This paper summarizes and proposes a framework for the health assessment of complex systems， which consists of four aspects： data acquisition， data processing， health assessment， and health prediction. At last， the prospect view of health assessment is presented， and a system-leveled open-source project， namely OpenHA （Open Health Assessment） is suggested.

Based on the application scenarios of swarm intelligence in the field of UAVs， the application of swarm intelligence methods in the field of UAVs was reviewed. First， the recent application status of UAVs was reviewed， and the principles of swarm intelligence algorithms and examples of UAV applications were introduced. Second， the application scenarios of swarm intelligence in UAVs were divided into four parts： swarm intelligence-based UAV wireless communication， swarm intelligence-based UAV ad hoc network， swarm intelligence-based UAV trajectory planning， and swarm intelligence-based UAV intelligent decision-making. The progress of relevant research work for each part is introduced separately. Finally， a brief discussion is conducted on the development trend of swarm intelligence for UAVs.

In order to solve the problem of speed planning in complex road environment （large curvature， low adhesion） meeting the constraints of safety and efficiency， a differential equation programming method based on vehicle dynamics was proposed. Firstly， the structural parameter expression of the limit velocity satisfying the lateral tire force constraint was derived in steady-state steering. Secondly， the spatial combination of the tire forces of the front wheel and the rear wheel is shown in the F-F diagram. And the implicit differential equation considering load transfer and driving mode factors was derived. The limit velocity along the path can be obtained by solving the differential equation. A method for calculating the limit speed based on discrete path information was given. Finally， a model prediction controller was designed and a co-simulation platform of CarSim and Simulink was built. The trajectory tracking simulation experiments were carried out with the planned limit speed on the continuous and discrete information paths. The results show that the proposed limit speed planning method can complete the trajectory tracking task as soon as possible in the complex road environment and control the- tire force within the range of stable friction circle.

To address the problem of unpredictable and variable errors when using wheel odometry for localization， a wheel odometry error prediction model based on Transformer neural network is developed to accurately predict the odometry error that accumulates and changes as the mileage increases， and to improve the accuracy of localization using wheel odometry under GPS occlusion. First， two models were established without and with the driving condition characteristics， then they were compared with the LSTM model under various driving conditions. The experimental results show that the Transformer-based wheel odometry error prediction model can accurately predict the odometry error with higher accuracy， stability and reliability than the LSTM model under both regular driving conditions and challenging driving conditions where it is difficult to measure the odometry signal accurately. At the same time， compared with the Transformer model without considering the driving condition characteristics， the Transformer model with considering the driving condition characteristics improve the performance in all evaluation indexes， which proves that considering the driving condition characteristics can effectively improve the prediction performance of the model.

In the traditional technology of lateral localization using lane markings identified by cameras， incorrect matching of lanes or their left and right boundary points causes large localization errors under driving conditions such as lane changing. A multi-indicator weighted evaluation map matching algorithm combined with lane change recognition method was proposed. Further， a lane left and right boundary point determination method was designed， and a camera-based bilateral boundary line lateral localization method was proposed based on accurate matching to lanes and their boundary points， which can improve the lateral localization accuracy relative to lanes， and then fuse the camera with GPS， IMU， wheel odometer， and lightweight lane level map to form a complete localization system. The experimental results show that the accuracy and stability are significantly improved compared with the traditional fusion localization method using cameras and lane markings. The localization system designed in this paper provides a low-cost and high-accuracy solution for autonomous driving localization.

Aiming at the problem of driver's expected trajectory prediction， a vehicle expected trajectory prediction model considering the information of surrounding vehicles was designed. Separate encoders of self vehicle and surrounding vehicles historical track information is established， and the encoded self vehicle historical track information is sent to the intention recognition module to identify the driver's intention. The self vehicle and surrounding vehicles coding information is processed through the attention mechanism， the result of intention recognition is used as the input of the decoder module， and the future position of the vehicle is output. Finally,the data set is used to train the model， and the effectiveness of the model is verified.

A deep reinforcement learning agent based on Deep Q-Network （DQN） algorithm was constructed to solve the problem that the state machine decision model cannot effectively deal with the rich context information and the influence of uncertain factors in the snow and ice environment. The motion planner was used to augment the agent， and the rule-based decision planning module and the deep reinforcement learning model were integrated together to build the DQN-planner model， so as to improve the convergence speed and driving ability of the reinforcement learning agent. Finally， the driving ability of DQN model and DQN-planner on ice and snow road with low adhesion coefficient is compared based on CARLA simulation platform， and the training process and verification results are analyzed respectively.

Aiming at the problem of intelligent vehicle's driving environment understanding and behavior decision-making in complex scenarios， an integrated moving horizon decision-making method for lane and speed of intelligent vehicle was proposed. To obtain the vehicle positions based on the centerlines of different lanes， the vehicle-road space reconstruction model in complex scenarios and the integrated model of lane and vehicle speed with non-integer longitudinal vehicle speed and integer lane number as control variables was established. The safety of future trajectories of the vehicle and oncoming vehicles from multiple directions was analyzed. The integrated decision-making of lane and speed of intelligent vehicle in complex scenarios was described as a mixed-integer nonlinear programming problem. In order to verify the effectiveness of the integrated moving horizon decision-making method of lane and speed， the joint simulation of the vehicle dynamics simulation software veDYNA and Simulink is carried out in the unprotected intersection scenario. The results show that the decision-making method can change lanes in the face of slow-moving or abruptly cut-in vehicles； for safety, intelligent vehicle can make decisions to slow down and give way, when it and the surrounding vehicles drive to the intersection area at the same time.

Aiming at the problems of computing power waste and repetitive testing in the current accelerate test methods which enhanced generating of hazardous scenarios， an accelerate test method based on hazardous boundary search for the automated driving system is proposed. Firstly， three types of inherent attributes of the scenarios are defined： scenario hazardous degree， exposure frequency and sensitivity， and the scenario test priority is proposed with the three inherent attributes which is treated as the basis of the division of parameter space and the search sequence of scenarios. Next， specific scenarios are extracted according to the scenario test priority； and the scenario test priority is updated according to the scenario hazard degree in test results by iterating the update process. Then， the support vector regression is used to fit the hazardous boundary of the experimental results. Finally， the Matlab/PreScan/CarSim co-simulation platform is built and a black-box automated driving algorithm is tested with the front vehicle cut-in scenario in virtual environment to verify the effectiveness of the proposed method. The results show that the proposed method can effectively search the hazardous boundary of the tested automated driving algorithm and improve the test efficiency.

Aiming at the problem that the vehicle is more prone to instability in ice and snow weather， the stable tracking of the vehicle to the reference path is studied under the condition of low adhesion and asymmetric road surface. At the same time， considering the cooperative sharing between the driver and the intelligent controller， the sharing mode and driving right allocation is discussed， so as to improve the tracking accuracy and steering stability of the vehicle under the complex road conditions of ice and snow road and ensure the driving experience of the driver. A vehicle model with asymmetric left and right wheels for snow and ice road surface is established. And the vehicle steering stability constraint under ice and snow pavement is determined as a part of the subsequent model predictive control solution. The human-vehicle sharing structure based on model prediction and fuzzy weight allocation strategy is established and the shared steering controller is designed. Simulink/ CarSim co-simulation verifies that the designed control system can effectively improve vehicle tracking accuracy and driving stability.

To solve the string stability problem of cooperative control of constrained vehicle platoon with predecessor-following topologies， the distributed parameterized model predictive platooning control algorithm was proposed. Firstly， the longitudinal kinematic model of vehicle platoons is established， which then is transformed into a linear state space model via feedback linearization. Secondly， the local controllers of the vehicle platoon are determined by solving the local optimization problems of vehicles constructed using the predictive trajectory information of neighborhood vehicles. To reduce the computational demand of solving the local optimal control problem， the incremental control input over the prediction horizon is parameterized as the form of the stair structure. Then an iterative algorithm of model predictive platooning controller is designed to satisfy the constraints of input/output and string ability. Furthermore， by using Lyapunov stability theorem and Moore-Penrose inverse matrix theory， the sufficient conditions are established to ensure asymptotic stability of the vehicle platoon as well as the analysis of the tracking performance. Finally， the effectiveness of the proposed control strategy is verified by comparison simulation with conventional platooning control algorithm.

Aiming at the problems of high model complexity of control system and the tire cornering stiffness uncertainty affecting the control effect caused by longitudinal and lateral coupling nonlinearity of intelligent vehicle， a longitudinal and lateral integrated feedback linearization control method based on Lyapunov stability theory and tire cornering stiffness estimation is proposed. Firstly， the longitudinal and lateral coupling model and the trajectory tracking error model was established. Secondly， the exact linearization condition of the system was determined， and the virtual control law was designed to ensure the stability of the system and the asymptotic convergence of the tracking error by using the Lyapunov stability analysis method. Then， the real-time estimation of the tire cornering stiffness was carried out. The simulation results based on CarSim/Simulink demonstrate that the proposed method can make the intelligent vehicle keep good trajectory tracking performance and stability in longitudinal and lateral coupling conditions.

Decision-making and motion planning are the crucial modules of autonomous vehicles. Focusing on the driving parameter relationship between decision making module and motion planning module， and the dynamic characteristics of new hybrid traffic flow， lane-changing model of autonomous vehicles was established. The decision-making process of lane-changing of autonomous vehicles was described by Stackelberg game theory， and the game cost function is quantified by negative exponential function. The decision results of the autonomous vehicle were taken as one of the inputs of the motion planning module. Polynomial was used to describe the lateral and longitudinal trajectories of the vehicle， and simulated annealing algorithm was used to find the optimal trajectory. The simulation results show that the proposed decision programming model can quickly generate safe， comfortable and feasible trajectory.

To identify drivers’ takeover behavior and intention in an autonomous driving environment， with the help of a driving simulator and an eye-tracking device， the simulation tests were conducted to let drivers takeover an autonomous vehicle ten times in five emergency situations on a two-way six-lane freeway of 18.95 km. Using the vehicle operation and visual attention data， the three common factors were extracted by the factor analysis， and the K-means clustering analysis was used to qualitatively identify the drivers' takeover behavior and intention. The factor analysis was respectively combined with the support vector machine and the long short-term memory neural network， then the two models were obtained to quantitatively identify drivers' takeover behavior and intention. Research results show that： drivers' takeover behavior is influenced by their longitudinal response， lateral response and visual attention； the clustering analysis can qualitatively describe the takeover behavior and intention of different types of drivers and reveal the potential driving safety risks； compared with the support vector machine， the long short-term memory neural network and the factor and support vector machine model， the factor and long short-term memory model is more effective in identifying drivers' takeover intention， with the best four performance indices of accuracy rate， recall rate， F₁-score and precision rate； and the common factors which are obtained using the factor analysis for data downscaling and effective information enrichment are helpful to improve the classification performance of the driving takeover intention recognition model. This study is helpful to identify drivers who are at a higher risk of takeover and to design some targeted driving assistance strategies.

A multi-mode interactive filtering algorithm that integrates Global Positioning System （GPS） and vehicle sensor information was proposed， so that the vehicle can adapt to a variety of driving conditions. The longitudinal and lateral motion model were built to describe the full coverage prediction of the expected driving position of the vehicle. The credibility evaluation of GPS signal was realized by data-driven neural network and vehicle motion model. The accurate estimation of process error and measurement error was realized by interactive filtering and probability updating between vehicle motion models. The positioning fusion algorithm was verified in the real vehicle testing environment. The results show that the proposed algorithm can improve the positioning accuracy of the multi-sensor positioning system and provide more accurate position estimation results in the case of unstable signals.

In allusion to the disadvantages of traditional RRT algorithm such as strong randomness， slow convergence， lack of continuity and being unable to meet the dynamic constraints of most robots， and Kinodynamic RRT^* algorithm which conforms to dynamic constraints but have a slow search speed， a DKRRT^*-APF algorithm besed on Kinodynamic RRT^* algorithm was proposed. Kinodynamic RRT^* node expansion was improved from unidirectional expansion to bidirectional expansion to speed up trajectory search efficiency； meanwhile， in the sampling stage of Kinodynamic RRT^*， the idea of map potential field of APF algorithm was introduced， and special gravity function and repulsion function were designed to make the target points and obstacles in the map provide some target guidance and obstacle avoidance effect for the sampling nodes， and shorten trajectory planning time. To verify the feasibility of this idea， the DKRRT^*-APF algorithm was tested by using the double integral dynamic unmanned vehicle model and the linearized unmanned quadrotor model. The simulation results show that the trajectory planning by DKRRT^*-APF algorithm conforms to the dynamic constraints of the robot model. When the average trajectory loss is the same， the average number of generated nodes and the average trajectory planning time of DKRRT^*-APF algorithm are effectively improved compared with KRRT*.

In order to accurately and reasonably predict the future trajectories of vehicles and understand the changes of surrounding traffic flow, a trajectory prediction method combined with high definition map in graph attention mode was proposed. The encoder-decoder framework based on LSTM network was designed, and the model structure with vehicle historical status and high-precision map information as input was established. A graph query mechanism combining local and global features of vehicles was proposed to output vehicle prediction trajectory. The results of experiments carried out on the nuScenes dataset show that the comprehensive prediction performance of our model is better than other state-of-the-art methods， such as Traj++， CoverNet， etc.， and it has good anti-interference.

A novel 3D object detection algorithm based on high-precision map enhancement （HME3D） was proposed by integrating high-precision map information into the backbone detection network. Specifically， the high-precision map feature extraction module （HFE） was constructed by combining traditional convolution and transformer to achieve efficient extraction of map features. In addition， the auxiliary supervision network （MEES） based on map edge enhancement was designed to improve the performance of the main 3D object detection task. Finally， the advantages of the proposed model in this paper are verified on the challenging nuScenes dataset， which improves the accuracy of the LiDAR baseline model by 2.81 mAP.

Aiming at the problem of actuator tracking error in the commonly used “planning-tracking”obstacle avoidance method for quadrotors， an obstacle avoidance position controller with obstacle constraints was proposed by adopting an integrated design scheme of control-anti-disturbance-obstacle avoidance. The Barrier Lyapunov function was used to constrain obstacle boundary. The target point becomes the balance point by introducing distance information between the aircraft and the target position， through which the problem of the unreachable target in traditional potential field methods was solved. For the problems of strong state coupling and inaccurate model building in the quadrotor control system， an observer-based model compensation control strategy is proposed and applied to the attitude control. The compensation function observer is used to estimate the model-bias and external disturbance， and then the estimated value is fed back compensated to the controller in real time to achieve the adaptive anti-disturbance control effect. Finally， the above algorithm is simulated and verified， and the results show that the observer-based model compensation control has better control effects than other control algorithms in transient performance， tracking expected response and anti-disturbance. The obstacle avoidance position controller does not need to consider the problem of tracking error. In terms of time in simulation， the integrated obstacle avoidance method is greatly shortened compared with traditional "planning-tracking" obstacle avoidance method， and static obstacle avoidance can be achieved by giving the starting and target positions.

A fault detection method for multirotor UAV formation in complex wind field flight environments was proposed to address the shortage of multirotor UAV formation fault detection research. Combining multirotor dynamic behavior and wind disturbance modelling， the state variables of each multirotor in the formation were estimated in real time based on the Kalman filter algorithm， and the fault features were extracted by means of probabilistic statistics. Then， an identification model was built for fault identification， and the final result was obtained for the identification of the faulty aircraft in the formation. In the simulation validation session， a complex wind field model with superimposed turbulent wind and colored gusts was built based on the RflySim flight simulation platform. In order to assess the efficiency of the approach， based on this platform， the propellers， accelerometers， GPS and batteries of the multirotor are injected into the aircraft， and flight simulations of the multirotor UAV formation are carried out.

In order to improve the efficiency of multi-unmanned aerial vehicle search， reduce the error impact caused by operation and communication delay， and improve the positioning accuracy in the actual search process： the man-machine manually searched for the frame to select the target and extracted the shared information of the target； then the other machine used YOLOv3 to automatically detect and locate the target according to the shared information； finally， a KCF-based delay error compensation algorithm was designed to estimate and compensate the coordinate error in the selection process. The experimental results show that the above method effectively improves the efficiency and accuracy of multi-machine search and positioning.

A model predictive control autonomous maintenance strategy based on risk evaluation function was proposed to address the remaining life shortening problem caused by actuator degradation during the operation of quadrotor unmanned aerial vehicle(UAV）. Firstly， the architecture of autonomous maintenance system was constructed through in-depth analysis of the control mechanism of the UAV. Secondly， based on the definition of the system failure threshold， the analytical solution of the remaining life distribution of the UAV under hovering condition was obtained， and the values of Q and R elements of the prediction weight matrix of the model were adaptively modified based on the risk evaluation results to realize autonomous maintenance， so as to achieve a better compromise between UAV performance and life. Finally， the simulation experiment results show that the proposed autonomous maintenance strategy can extend the remaining life of the UAV with implied actuator degradation by 616 min and maintain better airframe performance.

Due to the limitation of the flight environment of the quadrotor UAV and its own actuator， its attitude is often subject to various constraints. To ensure that the attitude always changes within the constraint range， the barrier Lyapunov function （BLF） was introduced to ensure that the attitude always varies within the constraint range to ensure that the system can achieve the preset performance. A compensation function observer based BLF attitude constrained backstepping control （CFO-BLF） scheme was proposed. Simulink simulation compares the CFO-BLF backstepping control， ESO-BLF backstepping control and PID control algorithms by testing the attitude tracking performance of the quadrotor UAV to verify the effectiveness and superiority of CFO-BLF.

Aiming at the problem that the online trajectory planning of high-speed unmanned aerial vehicle (UAV) in the ascending phase needs to realize online fast solution under multiple constraints, firstly, the motion and dynamics model of the vehicle was built, and the constraints faced by the trajectory planning were given. According to the constraints and flight characteristics, the action state space and reward evaluation function that meet the mission requirements were designed based on the near end strategy optimization (PPO) strategy gradient optimization. Secondly, based on the characteristics of strong time memory of the trajectory planning in the ascending phase of the aircraft, the short and long term memory network (LSTM) network structure was introduced on the basis of the traditional PPO algorithm, and the PPO-LSTM algorithm was used to solve the online trajectory planning problem in the ascending phase of the high-speed aircraft, and the model that can plan the optimal angle of attack strategy in real time according to the aircraft state was trained. Finally, the performance of the algorithm was verified by Monte Carlo simulation. The results show that the root-mean-square error of the terminal state of the algorithm in this paper is reduced by about 50% compared with the traditional PPO and particle swarm optimization, which fully proves the superiority and effectiveness of the proposed algorithm.

A distributed adaptive sliding mode time-varying formation control scheme based on neural network state observer was proposed to solve the problem of unmeasurable information of multi-quadrotor unmanned aerial vehicles （UAVs） system in directed switching communication network. Firstly， by designing a distributed time-varying formation control protocol， UAVS can communicate with each other only by the status information of their neighbors， which is free from the dependence on global information. Due to the underdrive characteristic， a position-assisted controller was designed to solve the target information of two attitude angles and control thrust. At the same time， a neural network state observer was designed to observe the unmeasurable information of the system， and the observed values were feedback to the adaptive sliding mode controller in real time， which improves the robustness of the UAV system. The quadrotor UAV formation system is analyzed by Lyapunov's theorem， and it proved that the multi-UAV formation error is bounded and converges to near zero. The simulation results show that the proposed control method can realize the time-varying formation control of the multi-UAV system and verify the validity of the theoretical results.

In order to solve the attitude control problem of the coaxial tilt-rotor unmanned aerial vehicle which is influenced by external disturbance， a finite-time sliding mode controller （FTSMC） was designed to improve the tracking performance of attitude control and the robustness under uncertain disturbance. The Newton-Euler method was used to establish the attitude dynamics model of the UAV， the power distribution was adjusted to solve the over-actuated problem of the UAV， the stability of the FTSMC was verified by using Lyapunov theory， and finally， the simulation test was utilized in the Matlab/SimMechanics platform to verify the performance of the proposed controller.

Aiming at the problem that it is difficult to detect UAV in real time in public safety areas with complex flight backgrounds such as parks and playgrounds， the YOLOv5-Unmanned aerial vehicle object detection algorithm of anti-unmanned aerial vehicle system based on attention mechanism and the fusion of scale adaptive features is proposed. Firstly， the unmanned aerial vehicle data set is built by using fusion of selfie images and public data set DUT-Anti-Unmanned aerial vehicle. Secondly， the anchors is redesigned by using k-means method. Thirdly， the scale adaptive feature fusion module is designed. Then， the CIoU loss function is used as the positioning loss function of the YOLOv5s algorithm. Finally， the coordinate attention module is introduced in the backbone network to guide the network to pay attention to the channel and spatial position information of the unmanned aerial vehicle. The improved algorithm and the baseline algorithm are compared with the established data set， and the experimental results showed that the improved algorithm improved by 6.1%， 5.8%， and 5.2% in terms of precision rate， recall rate， and the average accuracy （mAP@0.5）， and the detection speed was 39 frame/s. On the dataset of this paper， compared with the YOLOv5m， YOLOv5l， and YOLOX object detection algorithms， the average accuracy （mAP@0.5） is 4.4%， 3.6%， and 1.3% higher than the comparison algorithms， indicating the effectiveness of the improved algorithm. A comparative experiment was conducted on the public data set VisDrone2019， the average accuracy （mAP@0.5） of the improved algorithm YOLOv5-Unmanned aerial vehicle for all categories is higher than the original algorithm YOLOv5s， and the detection effect of small objects in complex backgrounds is better， respectively.

In order to enable quadruped robot to independently adjust its gait planning parameters according to change of terrain without terrain sensing system， a static gait planning method based on GRU （Gate Recurrent Unit） model was proposed. Firstly， a method of generating continuous rectangular trajectory for swing foot was presented， which can ensure swing foot of robot move smoothly to target landing point on unknown terrain. Then， a planning method which can change the trajectory of body by adjusting parameters was proposed. Finally， GRU model and step times of each swing foot are used to realize prediction of the adjustable parameters in body trajectory planning， so that quadruped robot can generate a motion adapted to the change of terrain， and energy consumption and self-stability were considered in motion planning. The experimental results show that the proposed method is correct and effective.

Aiming at the problem of UAV secrecy communication and ensuring safety as well energy saving during flight in the wireless network scenario， a multi-objective optimization scheme was proposed. The scheme mainly includes UAV transmission model， UAV energy consumption model and environmental constraints model， and further constructs the multi-objective optimization model of UAV scheduling and path optimization problem （USPOP）， which optimizes average communication secrecy rate of UAV wireless communication， UAV hovering energy consumption and UAV flight energy consumption. Then， a non-dominated sorting genetic algorithm III （NSGA-III） with discrete normal distribution initialization， differential mechanism， genetic mechanism and avoiding obstacles operator （NDGA-NSGA-III） is proposed to solve USPOP. Simulation results show that the proposed algorithm can effectively solve the constructed optimization problem， and the convergence effect is better than other comparison algorithms.

For circumnavigation control of a multi-UAV system with multiple static targets was proposed bearing measurement based target localization and circumnavigation control algorithms to ensure that the multi-UAV system can realize the target localization and circumnavigation in fixed time. First， A bearing measurement based fixed-time target localization algorithm was designed for the leader-UAV to compute the position of each target，and a fixed-time observer was proposed for each follower-UAV to estimate the targets' center and the position of the leader-UAV. Then， a distributed fixed-time circumnavigation control algorithm was designed for the multi-UAV system to ensure the system realize a uniformly distributed circumnavigation of the multiple targets. Finally， numerical simulation experiment verifies the effectiveness of the proposed circumnavigation control algorithm.

Aiming at the problem that ground points in point cloud data can affect the precision and speed of environment perception， an accurate segmentation method of ground point cloud based on plane fitting was proposed. Firstly， the scene point cloud was divided into several areas based on the projection distance. Secondly， according to the average height in the area， the divided ground points and the normal vector direction of the ground plane， the ground plane fitting point was determined， and the ground plane was fitted. Finally， according the distance from the point to the ground plane to achieve ground segmentation. Using the KITTI dataset and the collected point cloud data to compare the proposed algorithm with four algorithms： RANSAC、GPF、R-GPF and PatchWork， verify the effectiveness of area division， fitting point screening and ground plane normal vector direction screening for ground segmentation. The experimental results show that after the area division， the far-distance sparse ground can be divided； after the fitting point screening， the ground segmentation accuracy reaches 0.9417 under the condition of low iteration. After screening the normal vector direction of the ground plane， fitting the wall to the ground was avoided. The proposed method is better than the four compared algorithms in terms of F₁ score， recall rate and accuracy rate， and the speed can reach 42.78 Hz， which can divide the ground accurately and quickly.