Study On Detection And Tracking Algorithm For Traffic Object Based On Video Image Sequence

At present, the detection and tracking of traffic object which is based on video sequencesis an important research orientation in the design of intelligent transportation system. It ismainly used in automatic vehicle driving, intersection detection, detection and recognition oftraffic sign, detection of traffic flow, surveillance for traffic safety and traffic management.The purpose of the object detection and tracking which is based Video designed is to makethe computer to be able to have their own perception of the environment capacity as humansand provide important data for the follow-up of the behavior analysis and understanding. Thepaper doesexploration and research about some key issues of the traffic video object detectionand tracking. The major research aspects involved include:(1) On basis of Gaussian mixture background model, we Bayesian-priori theory for stateestimation, and the distribution of conjugate prior for estimating model parameter,then putforward a background subtraction algorithm based on an improved adaptive Gaussianmixture model. The algorithm converts posterior distribution calculating to prior distributioncalculating, for solving the complex computational problems brought in by posteriorprobability, and can update background by getting the posterior distribution in real time, itperforms better in eliminating the problems brought by the changes in light intensity, thesystem disturbance and other reasons.(2) For the shadow detection algorithm, we uses the normalized criterion to differentiatesbetween shadow point and the target point near, according to the sensitivity of the RGBchannel of the shadow region before and after the occlusion.And we integrate suspectedshadow that is obtained by the color characteristics of shadow with suspected shadow whichis got by direct threshold, to judges the ultimate suspected shadow.(3) Taking into account that the traditional Kalman, Mean shift, particle filter algorithmmay have some defects in dealing with complex scenes, we propose a new trackingalgorithm that is a combination of improve Kalman algorithm based on Maximum fuzzyentropy Gaussian clustering with Mean shift algorithm. This algorithm uses Maximum fuzzyentropy Gaussian clustering for fusing target detection point, reducing search for the pointswhich does not support the distribution, thus it can quickly find starting position of theforecast target in the current frame, and then use Mean shift algorithm to search the targets inthe neighborhood of the location. Experiments show that this improved algorithm caneffectively deal with problems which are brought about by occlusion by static object in complex environments and great changes of the object shape and background.(4) On the basis of the previous statement of detection and tracking principle, we useVisual Studio2005and Opencv library as a development language to design and develop anexperimental platform based on MFC. It realize three mainstream detection algorithms andsuggested tracking algorithm, and tests the rationality of the algorithm.