Accurate and fast obstacle detection method for automotive applications based on stereo vision
Yi-Chin Tsai, Kuan-Hung Chen, Yun Chen, Jih-Hsiang Cheng · 2018
This paper describes a simple yet efficient method of obstacle detection. Different from other methods, this study utilizes the coordinate models of and the depth map generated from stereo cameras to accurately detect possible obstacles. The proposed algorithm searches on the depth map along vertical direction for pixels having the same disparity. Having the same disparity indicates that these pixels are from the same object, and such object is an obstacle on the road. After implementation, the average processing time of the proposed obstacle detection algorithm for HD 720p image requires only 4.0 milliseconds (ms) on Intel Core i7 3.6 GHz processor, and 16.3 ms on an embedded system, i.e., the NVIDIA Jetson TX1. The detection performance based on stereo vision is more precise and faster compared with 2-D image object recognition. By directly comparing the purchasing price, the hardware cost to use stereo camera is also much lower than a RADAR or LiDAR system.