Study On The Key Techniques For Autobody Measuring-points Detection Based On Binocular Stereo Vision

The increasing competitive pressure from the market and the higher quality demand from the customer are driving industries to win the market and customers through the reliable quality by quality control strictly. The traditional management method of quality is called back testing, however, this method can't accommodate the demand of the new times because of the acceleration of manufacturing pace. What companies hope is that the manufacturing process can be monitored and controlled, and the rate of inferior products is closed to zero so that the cost can be lower.Autobody is the biggest assembly of a car and composed of large and thin complicated dimensional steel stampings. Meanwhile its structure and manufacturing craftwork is very complex, whether the manufacturing quality is good or not effects the whole car's performance. Therefore, it is necessary to detect some autobody characteristic points, and by the measuring points the whole autobody manufacturing process can be supervised and quality-tracked. The traditional detection methods are almost detection offline, but detection online timely can be available with the development of computer vision and some relative techniques. Under this big background, some studies on autobody measuring points detection by binocular stereo vision are done. The following is the main study works:The first paragraph is introduction. Detection online is necessary due to the current quality management methods in companies. And have a brief historical retrospect to autobody detection instruments development. After that computer vision is expounded from the three aspects: the fundamental theory, developing process and the applications on the automotives. At last, conclude the types of autobody measuring points.The second paragraph is calibration. First, introduce the mathematic model of binocular stereo vision and summarize the resoluble methods of the parameters in the model. In that the imaging model of a camera is complex and the aberration coefficient is difficult to make certain, the forward neural network is adopted to calibrate the two cameras. During the training process of network, some optimization works are done by the following methods: improving the input data accuracy of network, normalizing the data sets, self-adaptively choosing the number of hidden layer, and choosing the weight and threshold, the generalization of neural network is improved a lot. After compared with classical planar calibration method, this method is verified to get higher calibration accuracy.The third paragraph is edge detection. After summarizing the development epitome of subpixel edge detection, a mounting hole in a car fender are chosen to detect by analyzing the fender's function and detection necessity. The whole process of edge detection includes two steps: first, edge is extracted only in the zone of mounting hole in photos according to the actual situation of online detection. Second, interpolation method is adopted to locate edge accurately further, and the accuracy is up to subpixel level.The fourth paragraph is stereo matching. The development situation of stereo matching is summarized, and gray ladder is chosen as matching characteristic to exploit a binocular stereo matching method based on the gray ladder images according to the gray distribution of images. The algorithm can achieve an unique matching between edge points in the images by three constrain conditions. Next the exploited algorithm is deep analyzed, after that, another stereo matching method—SIFT characteristic algorithm is briefly introduced.The fifth paragraph is experiment results and error analysis. At first, get the spacial coordinates of mounting hole's edge points. And by using the knowledge about linear regression and spacial analytic geometory, the spatial circle is fitted through three steps including fitting the special plane about the circle, the plane's rotation and fitting the planar circle. The relative dimensions of the mounting hole are obtained. At last, some analysis about dimension error is done, too. The sixth paragraph is conclusion and prospect. Conclusion includes the main study work and innovation in this paper, the applicability and the weaknesses of the measurement system. And the future research works and future basic research are summarized from the two angles of engineering value and commercial value.Above all, under the guidance of data-driving-quality, the study on the autobody measuring points detection by binocular stereo vision is actually meaningful. Some helpful experiment and exploration on the key techniques are done in this paper.