Abstract: Computer vision provides general and abundant information for the environment and task description. Multiple view geometry can be used for the unified geometric modeling of visual perception and control tasks. In this talk, visual perception and control results of intelligent vehicles and robotics will be presented.
Visual perception provides the necessary feedback, such as the vehicle’s motion states and drivable road regions, for control systems. Since the 3D information might be lost and image noises exist in the imaging process, the effective pose estimation and motion identification of vehicles are challenging. Besides, intelligent vehicles are generally involved in complex scenarios. Therefore, it is difficult to robustly detect the drivable road space for safe vehicle maneuvers. Optimization and observer theories are applied to reconstruct the geometric information of the scene based on multiple view geometry. Then, real-time vehicle states and drivable road region can be identified effectively based on the reconstructed geometric information.
Visual control exploits the visual information for task descriptions and for controlling intelligent vehicles and robotics through appropriate visual feedback control laws. Since the depth information is lost in the imaging process of monocular cameras, there exist model uncertainties for the controller design. Moreover, the limited field of view of the camera and the physical non-holonomic constraints of intelligent vehicles also have great influences on the stability and robustness of the closed-loop system. Multiple view geometry is used for the geometric modeling and scaled pose estimation. Then, Lyapunov methods are applied to design stabilizing control laws in the presence of model uncertainties and multiple constraints.
