Research Of The Object Detection And Localization For A Mobile Robot Based On A Hybrid Vision

The rapid developments of computer vision and artificial intelligence have brought wide application for mobile robots. As the object detection and localization based on vision is now a hotspot in the study of mobile robots, it makes great sense to realize the object detection and localization effectively.In previous literatures, many robots utilize an ordinary camera such as a monocular camera or binocular camera to get the scene information. Such camera can take clear images, but has a limit field of view. This requirement is also characteristic for the task of scene reconstruction where multiple viewpoints must be selected to acquire a map of the scene which increases the complexity and reduces the efficiency of information gaining. An omni-directional camera with a 360 degrees’ horizontal viewing angle can contain complete information about the scene. To combine the advantage mentioned above, a mobile robot on which an ordinary camera and an omni-directional camera with a hyperbolic mirror are combined to build up a hybrid vision system. The system is used to explore the scene and the corresponding key technologies are analyzed, such as:the object detection, object localization and search path planning. The major works are listed as follows:(1) To integrate the image data of the two cameras, a virtual camera based calibration method is proposed to compute the extrinsic parameters of the hybrid vision system. Based on this, the 3D reconstruction method of the hybrid vision system is developed.(2) To the problem that the matching results of Scale Invariant Feature Transform are heavily relies on threshold, a matching method of two steps for acquiring corresponding points is proposed. The shortest distance from the matching point to the corresponding epipolar in the omni-directional image is calculated and used as a constraint condition which eliminates the mismatched point obtained by SIFT.(3) To the problem that model-based object detection method is low efficiency, a detection method based on the vision attention algorithm is put forward to detect the object from the omni-directional image, and object localization based on reconstruction is finished.(4) According to the position and orientation of object and field of view of the ordinary camera, the viewpoint of each object is calculated and the clear image is obtained.(5) Moving cost is the criteria to determine the detection order and colony algorithm is used to find the shortest path.