Abstract:

In order to overcome the influence of network latency on the cooperative perception accuracy and simultaneously improve the point cloud feature expression capability， a cooperative perception method based on point cloud spatio-temporal feature compensation network for intelligent connected vehicles is proposed. Firstly， the point-to-pillar feature extraction method is used to process the raw point cloud data， and the local neighborhood features of the laser points are then spliced with pillar feature maps. Secondly， the temporal latency compensation module based on the PredRNN algorithm is designed to predict the point cloud features of historical frames received from the surrounding connected vehicles， so as to achieve the synchronization of point cloud features from two vehicles. Thirdly， the spatial feature fusion compensation module is utilized to aggregate the inter-vehicle point cloud features， and multi-resolution features are fused through the bidirectional multi-scale feature pyramid network. The output includes vehicle target geometry size， heading angle and other information. Finally， the test results on the V2V4real dataset and the self-collected dataset demonstrate that the detection accuracy of the proposed method is superior to classical cooperative perception algorithms. Furthermore， it exhibits good adaptability to various latency cases and the inference process meets the real-time requirements.

Key words: intelligent connected vehicles, collaborative perception, spatio-temporal feature compensation, point cloud, feature fusion