JYOLO: Joint Point Cloud for Autonomous Driving 3D Object Detection

Abstract

The camera and lidar are significant sensors for automatic driving, they can provide adequate complementary information. However, 3D point cloud object detection suffers from complexity and low accuracy. In this paper, a Joint-YOLO fusion model is proposed. It provides a low-complexity joint fusion object detection framework. First, the dilated attention is designed to pay attention to the feature resolution of correlation and reduce the number of calculations. And secondly, parallel inverted residual is constructed to connect deep and rich semantic information with high-dimensional features. Finally, the model present an efficient joint fusion structure embedded with camera-lidar detector based 2D-3D bounding box geometric and semantic information for 3D point cloud object detection.