A biologically-inspired vision architecture for resource-constrained intelligent vehicles

摘要

The use of computer vision for assisting the driver dates back to first research projects in 1980 s, but only recently the progress in vision research and the increase in computational power have resulted in actual products. Although impressive from the robustness point of view, these systems are optimized for specific problems and at best perform reactive tasks like, e.g., lane keeping assistance. However, for a better understanding of generic traffic situations and for assisting the driver in the full range of his actions, integrated and more flexible approaches are needed. In this contribution we propose a vision system that in important aspects is inspired by the human visual system for organizing the different visual routines that need to be carried out. The presented system searches for biological motivation in case classical engineering-based approaches cannot do better or fail. Using a tunable visual attention system and state-of-the-art perception algorithms, the system is capable of analyzing the scenery for task-relevant information in order to provide the driver with assistance in dangerous situations. Our main research focus is on the design of general mechanisms (i.e., not domain or task-specific) that lead to a certain observable behavior without being explicitly designed for this behavior. Using this principle, we aim at developing easily extensible driver assistance systems. The system components are evaluated on a complex inner-city scene and on further real-world data. We demonstrate the performance of the integrated vision system in a construction site setup. A traffic jam within the construction site results in a dangerous situation that the system has to identify in order to warn the driver. Different from other systems the detection of the dangerous situation is based on the vision channel alone. Radar is only used to assign distance data to visually detected objects. The contribution represents an important intermediate stage for future, more cognitive driver assistance systems.