A Novel Object Detection and Localization Approach via Combining Vision with Lidar Sensor

Abstract

This paper addresses the problem of object detecting and localization with a lack of global vision object information during the movement of a robot. A novel two components vision-based scheme is proposed in this paper. One is a light-weight convolution neural network (CNN) which is used to realize the detection of an object and its training model parameters are decreased to apply to the embedded system. The other is the lidar fusion structure which provides extra information of an object to improve the relative-localization capability of the robot. Experiments are designed to verify the effectiveness of this scheme. According to the results, compared with the existing method such as Yolov4-tiny, our scheme prominently increases the FPS of object detection by 60% and the two components enhance the robustness than the conventional method.