Abstract: Aiming at the problems of small-object recognition difficulty, sparse point clouds at long distances, and insufficient multimodal feature fusion in three-dimensional object detection for autonomous driving scenarios, we proposed an improved algorithm based on a multimodal three-dimensional object detection framework. The algorithm enhanced the ability  to preserve foreground information and suppress background noise interference by constructing a class and centroid-aware foreground point sampling strategy. By introducing  a dynamic convolutional image feature extraction mechanism,  the quality of image feature representation was improved. By designing a multi-stage interactive feature attention fusion module,  the deep collaborative modeling ability between point cloud features and image features was improved. Experimental results on a public dataset show that the proposed method achieves average detection accuracies of 83.49%, 46.98% and 68.28% for three types of objects: cars, pedestrians and cyclists, respectively, and outperforms current mainstream methods in overall performance. The proposed method can effectively improve the accuracy and robustness of three-dimensional object detection in complex traffic scenarios and has certain reference value  for promoting the development of autonomous driving environment perception technology.

Key words: computer vision, key point sampling, dynamic convolution, 3D object detection, multimodal fusion