Stereo Vision-Based Convolutional Networks for Object Detection in Driving Environments

Abstract

Deep learning has become the predominant paradigm in image recognition nowadays. Perception systems in vehicles can also benefit from the improved features provided by modern neural networks to increase the robustness of critical tasks such as obstacle avoidance. This work proposes a vision-based approach for on-road object detection which incorporates depth information from a stereo vision system within the framework of a state-of-art deep learning algorithm. Experiments performed on the KITTI benchmark show that the proposed approach results in significant improvements in the detection accuracy.