Moving Object Detection And Tracking Technologies In Complicated Traffic Monitoring Scenes

During moving objects detection and tracking there will be occurring unpredictable problems inintelligent transportation usually, because of various applications in various scenes, complex environment, andmoving objects’s the arbitrary, random, occlusion, light and so on. All of the uncertainty factors make movingobject detection and tracking to be difficult and challenging problem, and receive much attentionfrom academia.In the paper, the whole process is worked by a monocular camera, moving object detection andtracking are based on object feature extraction. According to the deformation and scale change of objects thepaper studied the object segment, recognition based on multifactorial and tracking in complex scenesimportantly, all of them formed a set of methods of object detection and tracking in image sequence.The main content of the paper as follows:1) A pixel-block codebook method is proposed based on quantum clustering analysis technology basedon the classic Codebook. In actual monitoring, the scenes are changing uncertainly, in order to extract theforeground precisely it is needed to the model background effectively to understand pixels changing and thecorrelation between of pixels in video sequence. The method divided the image into the same size blocks,learned and coded for those blocks, then using alternating update method to real-time update the code. Finally.Experiment shows that the foreground extracted by the method has less noise, and is clear. In addition, themethod is simple and adopts the method of alternating update; the method can extract the foreground real-timeand has the robustness.2) An adaptive cast shadow detection and elimination method is proposed. For detecting the shadow，information of color and local texture for foreground and background are compared. The HSI color space areused to detect shadow because it close to human visual characteristic, at the same time, according to localtexture changes of the foreground and background, the hamming distance of SILTP(Scale Invariant LocalTernary Pattern) is used to detect shadow too. The MLE(Maximum Likelihood Estimation) are used toestimate threshold occuring. MRF model is constructed to represent the dependencies between the label of apixel and the shadow models of its neighbors, and used to segment shadow and object. Experiment shows thatthe method has similar or superior performance comparing with other methods, and adapt to the environmentwith lighting change.3) An object recognition method is proposed which extracts histogram of oriented gradient based onGabor features images. In view of changing demand about the object pose and the light in the scene, object recognition and classification is the foundation of target tracking and behavior analysis effectively. Accordingto the characteristics of the Gabor wavelet, the method fused40Gabor wavelet features into one images on thescale and direction， and extracted HOG feature for it. Finally, Real Adaboost method is used to recognize theHOG features of objects. Experiments show that the method can reduce the error detection rate effectively, theamount of calculation is reduced effectively because of adapting the encoding for Gabor features in theprocess of image fusion.4) An object tracking method is proposed based on block sparse represented of object shape. The shapemodel of an object is built by a sparse linear combination of structured union of subspaces in a basis library，in order to realtime update template set the incremental learning are used, Then, a probabilistic observationmodel is built which based on the approximation error between the recovered image and the observed sample.The observation model is tracking by a particle filter framework formed by a stochastic affine motion model.Experiments show the method has a better effect comparing with IVT and L1method in dealing with suddenlight change，scale change and occlusion.