Research On Collision Prediction Of The Overhead Crane System Based On Stereo Vision

This thesis conducts a deep study on the emergency braking control for the crane system and presents a stereo vision-based collision prediction algorithm, which improves the safety and efficiency for crane operation. In addition, an experiment testbed is build to test the efficiency of the stereo vision-based collision prediction algorithm conveniently. In general, the main work of this thesis can be summarized as follows:(1) The experiment testbed of collision prediction based on stereo vision is constructed, and the corresponding software is developed for algorithm implementation. Specifically, the hardware component of the experiment testbed is composed of stereo vision cameras, a fixation bracket,a computer, a YanHua USB device and an optoelectronic trigger circuit. The software is programmed in the environment of Visual C++6.0 and Matlab 7.0.(2) A stereo calibration approach by the optimization of a global cost function is proposed. Based on Zhang's monocular camera calibration method and an appropriate distortion model, the proposed method can calibrate all parameters of stereo cameras efficiently, including intrinsic and extrinsic parameters of each camera, distortion coefficients, as well as the extrinsic parameters between cameras, by fusing a practical constraint to the global cost function. Compared with the conventional stereo calibration method, the proposed approach takes account of the fact that the extrinsic parameters between the two cameras are invariant.(3)An efficient edge-point matching algorithm with rotation-invariant features is presented. The proposed method successfully combines the SIFT algorithm with the conventional edge-point matching algorithm to improve the efficiency and reliability of the edge-point matching process. First, the contour edges of the sample image and the searching image are detected with the boundaries of the object. Then, according to epipolar constraint in stereo vision, the descriptor of the feature point is created by the local gray region at the point location, along the dominant orientation to the feature point. Subsequently, a candidate matching point is chosen in the light of redundant angle thresholds. The matching result is finally obtained by calculating the correlation between the descriptors of the feature point and the candidate matching point.(4)A multi-object detection and tracking and size estimation algorithm based on the Camshift (Continuously adaptive meanshift) algorithm is designed. The proposed algorithm utilizes background subtraction to detect the moving objects from the current frame image. After binary processing and morphology, the number and color distributions of the moving objects can be obtained to accomplish the tracking initialization. Meanwhile, the location as well as the size of the moving objects can be estimated by the tracking window information of the Camshift algorithm. Finally, the Kalman filtering technique is utilized to optimize the motion information.(5)A bounding box-based two-level collision prediction method is proposed in this thesis. First, the spheres bounding box is used to cull the scene in the project, only considering the obstacle within safe distance around the load. Second, the main task is to predict the possible collision between the load and the obstacles within the safe distance by utilizing the collision prediction algorithm which is based on the relative velocity. According to the chosen bounding box model, different collision prediction algorithms are put forward. For spheres bounding box, the thesis proposes a nearest point collision prediction algorithm based on the relative velocity. Then, for OBB (Oriented bounding box), the paper further presents a point-to-plane collision prediction algorithm, which is also on basis of the relative velocity.