Research And Implementation Of A Vehicle Image Segmentation Algorithm In Bad Weather

In the on-board automotive target detection system, image segmentation can process image sequence which were obtained by on-board camera, and then output the external rectangular area of vehicles, that is region of interest (ROI), then send the information to the next Object Recognition Module to do the further judgment (vehicle or not). On the basis of previous studies, an improved vehicle segmentation algorithm in bad weather has been proposed and implemented in this thesis, and the vehicles were located behind the camera.By analyzing and summarizing the advantage and disadvantage of existing vehicle segmentation algorithm, this paper has proposed a new vehicle segmentation algorithm in bad weather. Via analyzing large amounts of vehicle image feature in bad weather and studying different principles of image enhancement method, the thesis have summarized the general character of vehicle image in bad weather and have proposed more adaptive image enhancement method for follow algorithm. For generating more distinct vehicle edge, making the generated vehicle edge more suited to follow algorithm implementation, based-on human visual specialties, this paper proposed coarse-grain binary image generating algorithm. That improved the algorithm performance to a great extent while the recognition has been increased. In addition, the important shadow image, vertical edge image generating algorithm and the strong robustness vehicle segmentation and vehicle location algorithm have been proposed. We also proposed a good threshold auto adaptive method. In the end, the algorithm performance, recognition rate and error recognition rate contrast data between improved algorithm and existing algorithm has been listed.The algorithms in this thesis have successfully fulfilled and applied in vehicle recognition engine from AAC (Advanced Automotive Technology Research Center), which include all kinds of road conditions and climate and so on and then vehicles in video from onboard camera are accurately located. Experimental results verify the robustness and applicability of the algorithms. Through the actual experiment on the real road, the algorithm has achieved expected results in recognition rate and real-time fields.