The Research Of Visual Pedestrian Detection,recognition And Tracking Techniques Under Moving Background

In recent years,the pedestrian detection and tracking techniques based on computer vision have achieved significant progress,there’re enormous applications appearing in such as visual surveillance,object behaviour analysis,robot control,human-computer interface,intelligent traffic,etc. Also,there are some hot developing applications including computer-aided automotive driving,self-driving automotive,driveless car,etc. Faced with the more and more complicated application scenes,especially the fast-changing background, the object recognition techniques purely based on computer vision are caught in more and more severe challenges. So far,we still haven’t got a robust,precise and fast object detection and tracking technique.Take pedestrian detection as an example,the pedestrians have vast in-class variances,especially the remarkable changes in clothing,lighting,and the inter-influence among pedestrians,and with the environment, interferences from human-like objects. As a consequence,the various detection algorithms still can’t meet the application demands in the real complicated environment. As to the more widely-applied object tracking techniques, research under static background has almost been mature,but for the moving background,the research is far not enough. The main contribution of this thesis includes the following aspects:1) Researched the fast pedestrian detection algorithm on single frame.As the pedestrian detection algorithms based on vision advance, it becomes more difficult to further improve the pedestrian detector’s performance. At the same time, complicated algorithms bring about more computation complexity, which severely degrade the detection system’s time efficiency. We consider to improve the algorithms’ robustness by minor changes without any time efficiency loss under the single-frame condition.This thesis mainly researches the fast pedestrian algorithm based on Adaboost +Chn Ftrs,where we find some more excellent detectors by choosing different feature combinations. In addition,we use the feature LUT method in Adaboost learning,which significantly improves the training speed, and increases the training data’s sample capacity a bit.2) Done the comparison of the tracking performance for different online classification algorithms and object feature model,promoted the online object color model, structure model based on the Fern classification method, which achievedbetter tracking results under experiments.Consider the common feature types,like color feature,struct feature,model,etc,which can be combined with different online classification algorithms to build object’s online classification model,among them the color feature uses the SPT algorithm based on superpixel color histogram, the structure feature uses the CT algorithm based on compressive feature,as to CT, we also take advantage of the compressive feature extracted from multi-channel images to refine the algorithm,the model matching uses the nearest neighbour classifier(NNClassifier) and the TLD algorithm,we point out the the adaptation scenes and drawbacks of all kinds of object feature’s online classification models. Furthermore, we promote the Fern algorithm combined with the color superpixel feature,and the multi-channel spatial minor rectangular feature, which realize better tracking results, under all kinds of complicated scenes like occlusion, lighting change, object posture change,object size change, and dynamic background.3) Considering the general cases for object tracking, to better adapt to the object size change,we adopted the particle swarm optimized particle filter(PSOPF) as the tracking filter algorithm. The experiments demonstrated that the particle swarm optimization could remarkably improve the performance for the particle filter, and also largely improve the overall tracking performance.4) At last, we promoted a relative real-time and robust tracking algorithm for object like pedestrians under moving background and complicated scenes, which could adapt to all kinds of tracking scenes including object posture and size variance,lighting change, occlusion, inteference,etc.The algorithm adopts the online object classification model based on multi-feature, and combined with the particle swarm optimized particle filter(PSOPF) method to do tracking. The multi-feature model includes the Fern color model, Fern structure model,and the CT adaptive structure model, all these feature models have strong compensations for each other, and the combined one can achieve quite excellent tracking results. In addition,in this algorithm we introduce in the nearest neighbour classifier(NNClassifier) based on direct model match as an inspective model,which can help to resist the tracking drift due to model’s slow adaptation without any computation complexity gain. We also promote a method to deal with the lighting-variance problem(among the multi-feature model,the color model is very sensitive the lighting changes), which greatly contributes to the objecttracking algorithm under complicated lighting-change scenes.