3D Lidar Target Detection Method at the Edge for the Cloud Continuum

Abstract

In the internet of things, machine learning at the edge of cloud continuum is developing rapidly, providing more convenient services for design developers. The paper proposes a lidar target detection method based on scene density-awareness network for cloud continuum. The density-awareness network architecture is designed, and the context column feature network is proposed. The BEV density attention feature network is designed by cascading the density feature map with the spatial attention mechanism, and then connected with the BEV column feature network to generate the ablation BEV map. Multi-head detector is designed to regress the object center point, scale size and direction, and loss function is used for active supervision. The experiment is conducted on Alibaba Cloud services. On the validation dataset of KITTI, the 3D objects and BEV objects are detected and evaluated for three types of objects. The results show that most of the AP values of the density-awareness model proposed in this paper are higher than other methods, and the detection time is 0.09 s, which can meet the requirements of high accuracy and real-time of vehicle-borne lidar target detection.