Accuracy Requirements of Camera-Based Depth Estimation for Urban Automated Driving

For autonomous driving in urban areas higher accuracy requirements for localization of surrounding traffic participants become apparent. The use of cost-efficient camera sensors shows potential for a performant depth estimation and can supplement perception systems to achieve redundancy. Current research focuses on improving the algorithms towards better performance whereas the application-oriented analysis of present estimation errors in relation to urban traffic scenarios is often neglected. Based on stereo and mono camera images, a benchmark analysis of rule- and deep learning-based depth estimation approaches is conducted in this work. The error-prone estimation results are then analyzed against braking distances of urban traffic scenarios simulated by a two-track model to analyze the criticality of different depth estimation approaches. The application-oriented evaluation shows that current approaches could already be used in real automated driving systems and enable the definition of requirements.