Research On The Algorithm Of Object Detection And Tracking Based On Monocular Vision

The intelligent video surveillance system requires more and more to the performance of intelligence with the rapid development of security monitoring industry during the recent years.The intelligent video surveillance is the system that autonomously detects the objects and tracks them,or which is the intelligent activity analysis system in a surveillance screen under the condition of without human intervention.The intelligent video surveillance mainly relates to the fields of artificial intelligence,image processing and pattern recognition etc..Object detection and tracking is not only the important part of the intelligent video surveillance system,but also the most basic one for whose performance will directly affects the quality of the whole system.It is still a challenging research topic while much process has made in the past decades.It may occur drifting and even tracking failure due to the factor of clutter background,scale variation and illumination variation and so on while in the real environment.In this paper,some researches have done about the common detection and tracking algorithms at home and abroad in view of the phenomenon mentioned above,and further improvements are made on this basis.The main attribution in this paper can be summarized as follows:(1)In the detection process,the traditional algorithm combined the HOG(Histogram of Oriented Gradient)and SVM(Support Vector Machine)is improved for person detection.While we often meet the false alarms due to the drawbacks of the traditional algorithm.In this paper,the confidence for each detected candidate is computed,and then a threshold is used to eliminate the false alarms.The detection process has provided a basis for following object tracking;(2)In the single object tracking process,an online object tracking algorithm using sparse and structural model is proposed.The method exploits local and structural feature to represent the target with sparsity.Furthermore,the paper uses the L2-pooling method to pool the sparse coefficient,and the target can be located more accurately and the occlusion problem also can be solved with this strategy.In addition,an adaptive updating strategy based on increment subspace learning and sparse representation is developed.The proposed adaptive updating strategy not only helps weaken the influence of illumination,but also reduces the possibility of drifting.Numerous experiments demonstrate that the proposed algorithm performs more robustly and effectively against several state-of-the-art algorithms.(3)In the multi-objects tracking process,a multi-objects tracking algorithm by collaborative multi-feature based on kalman filter is proposed.For each detected target,every target is tracked independently to avoid the high computational complexity of the joint probability with increasing number of targets.Furthermore,a collaborative model is constructed:color feature-based appearance model,texture feature-based appearance model and a spatial distance information model;Finally,for data association,a likelihood function is formulated to choose best matches for the targets and the candidates.The experimental results show that the proposed algorithm performs well while the targets are occluded,in/out of the surveillance of field.