3D Object Detection of Cars and Pedestrians by Deep Neural Networks from Unit-Sharing One-Shot NAS

This work adopts camera, lidar and radar sensors to realize 3D object detection of cars and humans in varied and complicated road scenes. Based on the Radar Voxel Fusion Net (RVF-Net) consisting of voxel feature extraction module, feature aggregation module, and detection head module, we extend layers of the voxel feature extraction module with dimensions at a consistent scaling ratio, and increase a prediction head in the detection head module. According to this modified RVF-Net, the depth of the voxel feature extraction module is searched by the proposed unit-sharing one-shot Neural Architecture Search (NAS) scheme where there are five choices of bypass, and one to four copy units for each layer. Only four copy units with conditional links yield five choices for structure searches. Such an approach can minimize the search cost and parameter amount. By using the forward-looking sensing data of nuScenes, the network from the proposed NAS, and the modified RVF-Net can yield superior Average Precisions (AP), more than 2.3% AP increase, than the conventional ones in detecting cars and pedestrians simultaneously. Therefore, the proposed network can effectively accomplish multi-object recognition tasks for self-driving applications.