Self-Supervised Stereo Depth Estimation: Depth estimation in multiple environments through an adaptive CNN and IR light

Abstract

We have developed a complete depth sensor unit with a self-supervised neural network and stereo camera. The sensor is both adaptive during usage and can work in dark and low light environments with aid from IR spotlights. Disparity estimation via stereo cameras has shown great performance in combination with neural networks during recent years. The reason is because deep learning reduces the computational effort considerably compared to previous methods. However, the existing deep learning methods do not evaluate the depth measurements but rather the disparity estimation accuracy on available benchmark datasets. In difference to earlier work, this system has been evaluated with respect to depth measurement accuracy and suitable evaluation metrics have been developed. If the stereo camera is to be used as a reliable depth sensor the depth estimation quality needs to be ensured. From the thesis contributions a high-functional depth sensor unit can be developed with potential to surpass other sensors with respect to the amount of data obtained per second.