Real-time Object Detection Based on RGB-D Image

Abstract

Object detection for environment perception is one of the most critical steps for autonomous driving and driving assistant systems. Currently, considering the size of detection model, there is still a need for further improvement to satisfy the requirement of real-time, high-accuracy in object detection approaches for embedded on-board devices. In this paper, we proposed a novel method that taking advantage of fully convolutional neural network and RGB-D image to enhance the speed of detection and simplify the model by reducing its size. Specifically, we utilized a simulator to obtain a lot of binocular and supplementary RGB-D images for training. The proposed method was evaluated on the KITTI dataset and can achieve an accuracy of 90% for easy task. Moreover, the simulation results demonstrated the efficiency of detection get a speed up to 180 FPS when running on the GPU.