Vehicle-infrastructure cooperative 3D target detection based on Feature Prediction Atrous Spatial Pyramid Pooling Net

Due to wireless communication delays in actual Vehicle-Infrastructure Cooperation (VIC) data transmission, the misalignment in time and space of the data during fusion makes VIC sensor data fusion more challenging compared to on-board data fusion. To address the issues caused by wireless communication delays, this paper presents a VIC3D Detection algorithm based on Feature Prediction Atrous Spatial Pyramid Pooling Net. Firstly, based on the mid-fusion framework of traditional VIC3D Detection, the previous point cloud data have been integrated into the infrastructure. Then, by utilizing the current timestamp characteristics and first-order derivatives of roadside infrastructure, the feature prediction model proposed in this article can predict the timestamp features on the vehicle side. Secondly, a new VIC3D Detection atrous spatial pyramid pool with a single-scale feature extraction network and data transmission model is proposed; therefore, increasing the receiving field and reducing data transmission costs has an obvious effect. Finally, the experimental results compared with the non-fusion PointPillars method indicate that mAP@BEV under the delay conditions of 100 ms and 200 ms, the performance (IOU = 0.5) was improved by 12.05% and 11.82%, respectively, with the transmission cost not exceeding 1/17th of the original data transmission cost.