Research On Key Technologies Of Mixed Traffic Video Detection

Nowadays, many Chinese cities are facing severe problems of mixed traffic, whichneed to be resolved through the intelligent traffic. The extraction of the mixed trafficparameters with efficiency and accuracy is the basis of the management and the control ofintelligent traffic. The conventional detector is difficult to meet the needs of the parameterextraction of mixed traffic and it is urgent to develop one mixed traffic detection system toextract the parameter effectively. Therefore the studies on the mixed traffic video detectiontechnology as well as the developments of corresponding system have important significancefor improving the real-time performance and accuracy of urban traffic control that play animportant role in traffic management and traffic safety. At present there are many limitationson the development and realization of mixed traffic video detection system due to theinsufficiency of the researches on the key technology of mixed traffic video detection. Inview of meeting the technology and the practice demands of the mixed traffic videodetection, this dissertation studied and discussed a series of key techniques for mixed trafficvideo detection, including the universal calibration algorithm, the extraction andcompensation of the foreground information of mixed traffic, rapid classification andidentification of mixed traffic information, multiple targets tracking and occlusion handlingin mixed traffic information as well as system integration of mixed traffic video detection system. The main research works completed and important research results achieved in thisdissertation are summarized as follows:(1) Camera Calibration Algorithm through Pixel-angle ReflectionThe camera calibration methods have been widely studied, but the conventionalmethods still have some problems on rapidity and adaptability. On the basis of previousstudies, this dissertation proposed a camera calibration algorithm based on the pixel-anglereflection. According to the pinhole imaging linear model, the nonlinear relationshipsbetween the pixel-angles as well as the pixel-angle and real coordinates were considered, andthe calibration was optimized according to the actual mixed traffic video detection. Besides,the camera parameters such as altitude and inclination were adopted to expand theapplication scope of the proposed calibration algorithm. The proposed algorithm wasverified by contrast experiments, and the results shown that the algorithm has the advantageof high detection accuracy, wide range of adaptability, good flexibility and robustness, andthe well prospect of engineering application.(2) Foreground Information Extracting and Compensating Algorithm Based onthe Mixed Traffic History informationConsidering the complexity of the mixed traffic scene, the historical information of themixed traffic was introduced to improve the conventional edge detection algorithm by usingthe PWH (Partition Weighted Histogram)-based background model to acquire thebackground image and background edge. In the improved algorithm, the current edge information, background edge information and historical information are compared, and thenthe foreground edge information are acquired by filtering the background edge informationusing inference and secondary analysis; the obtained foreground edge information wascompensated and rectified using seeds growing algorithm based on the integration of thedifferential information; and the foreground edge information can be used for mixed trafficclassification and identification are obtained. Finally, with the foreground extractionalgorithm, the shadow processing and night detection were achieved.(3) Fast Classification and Identification Algorithm for Mixed Traffic Based onExtreme Learning MachineCommonly, the algorithm of mixed traffic video detection has the problems of lowidentification precision due to the lack of effective classification feature, and the problems ofpoor real-time performance due to the computational complexity of the conventionalintelligent algorithm. Considering these problems, this dissertation proposed a rapididentification and classification algorithm combined with the Extreme Learning Machine(ELM) on the basis of the previous studies. The proposed algorithm takes the edge contoureccentricity vector as the classification feature, solves the dimensional unification problemof edge contour eccentricity vector using angle-interval sampling method, and improves theproblem of low identification precision of the conventional algorithm. By introducing theminimax eccentricity ratio into the characteristic expression of the mixed traffic, theproposed algorithm overcomes the problem of low distinction degree between the features of pedestrian and bicycle. Moreover, the proposed algorithm can rapidly identify and classifythe mixed traffic information based on the feature selection and expression combined withELM classification mechanism, overcome the time-consuming problems using intelligentalgorithms such as Support Vector Machine (SVM), and thus achieve the accurate andefficient identification on mixed traffic information.(4) Multi-Targets Tracking and Occlusion Handling MechanismThis dissertation proposed a method for solving the problem of multi-targets trackingand occlusion identifying, which based on the Kalman filter and Principal ComponentsAnalysis applied to SIFT (PCA-SIFT). For the handling the occlusion, the Kalman filteringis combined with the real-world parameters to achieve the multi-target prediction and todetermine whether the occlusion is happened. For the multi-targets tracking, different trackstrategy is adopted according to the situation of the occlusion. For the cases of un-occlusion,the real-world coordinates constraint based Kalman filtering is used to ensure tracking speed;and for the cases of occlusion, the prediction information is used to lock the target area anddecrease the retrieval range, and then the PCA-SIFT algorithm is adopted to achieve quickmatch and complete the occlusion restoration. Consequently, the multiple motion targetsprediction and tracking can be achieved, and the rapidity and accuracy of the tracking can beensured. The proposed detection system was verified using the actual traffic video. Theexperimental result shows that the system is able to complete the mixed traffic videodetection task to extract a variety of mixed traffic parameters accurately. At the end of this dissertation, the research works are summarized including the mainresearch achievements, conclusions and innovative points. Furthermore, the further worksare proposed based on the deficiencies and limitations of this study.