Deep learning in low-power stereo vision accelerator for automotive

Various forms of Convolutional Neural Network (CNN) architectures are used as Deep Learning (DL) tools for learning the similarity measure on video patches in order to run the stereo matching algorithm - the most computationally intensive stage of the pipeline for the stereo vision function used in designing an autonomous car. We propose a hybrid system implementation of the algorithm for real-time, low-power and high-temperature environment. The accelerator part of our system is a programmable many-core system with a Map-Reduce Architecture. The paper describes and evaluates the proposed accelerator for different versions of the stereo matching algorithm.