A Neural Network Enhanced Stereo Vision Obstacle Detection and Avoidance System for Unmanned Ground Vehicle
Fanjun Liu, Binggang Cao · 2013
This paper presents a neural network enhanced stereo vision obstacle detection and avoidance system for unmanned ground vehicle.We build a neural network to learn the mapping for the left image to the right image under the assumption of a flat road.Using the trained neural network we map the left image to the right directly and we get the left remapped image.So obstacles can be detected using correlation values between the right image and the left remapped image.With detection result the system tells the unmanned vehicle how to avoid obstacles.Our system does not require intrinsic calibration of stereo cameras and it does not perform the two IPM transforms.With neural network's parallel processing our system reduces the computation expense and increases the real-time performance.Experimental results show that our system is practicable.