Visual Object Detection Based LiDAR Point Cloud Classification

Abstract

The environmental perception plays a pivotal role in autonomous driving tasks and demands robustness in cluttered dynamic environments such as complex urban scenarios. LiDAR being one of the popular perceptual sensor, suffers with large number of inaccurate object detections. This work is an extension of an ongoing research on multiple object detection and tracking. Where, Neural Network based approach is considered for visual detection to aid the LiDAR point cloud processing, and to address the inherent shortcoming of the sensor. It is inferred that the proposed framework would perform in real-time on an embedded platform. In addition, the separate processing of visual and LiDAR sensor data will enable switching to a light weight LiDAR only setup in runtime when required.