Multi-Object Detection in 3D Point Cloud's Range Image Using Deep-Learning Technique

In recent years, LIDARs (Light Detection and Ranging) have gained a lot of insight into various fields such as agriculture, astronomy, robotics, autonomous driving, forestry, etc. Point Clouds obtained through LIDAR have gained attention in the field of computer vision, robotics, and autonomous driving. In computer vision, there are several high level tasks namely classification, detection, tracking, segmentation and registration. Object detection in 3D Point Clouds is one of the challenges faced in the field of computer vision, which is one of the high level tasks. Due to the unordered nature of point cloud data, detecting objects by designing deep neural networks is not a straightforward task. This paper aims to provide a solution to detect the objects in a 3D point cloud data by transforming into 2D range image obtained from LIDAR using the improved YOLOV5 (You Only Look Once- Version5) approach by embedding the coordinate attention module. The improved YOLOV5 is used to detect objects in an image and conducted experiments on different sizes of custom 3D point cloud datasets with and without augmentation. The experimental results showed how the precision, recall and mean average precision (mAP) metrics improve as the dataset size increases. Also, the comparative results with YOLOV5 yielded better results on our data.