Research On Moving Object Detection And Tracking Algorithms Based On Spatial Temporal Slice Method

Object detection and tracking is a key problem in computer vision, and it is thebasis of follow-up treatment, such as visual scene analysis and semantic analysis.The technique of object detection and tracking has a wide range of applications inintelligent video surveillance, traffic, human-computer interaction, visual navigationof robots, virtual reality, medical image processing, national defense, etc.Traditional methods employ single frame or few frames of images to detect andtracking objects. The performances are affected by scene changing, object pose andscale variation, and object occlusion or missing, which may cause detection ortracking failures. Moreover, without the aids of object history information, themethods have no correcting strategies under the object missing or tracking errors.To solve these problems, we propose a new data space to detect and trackobjects. By introducing the spatial temporal slice method, the original objectdetection and tracking problems in the XY image space are transformed into theproblems in the XT data space. The XT data space infers the dynamic changinginformation of the X image dimension along the T time dimension. Thus, featureswhich are directly extracted from the XT space may include the dynamic motion andtrajectory information of the object. Besides, in order to solve the lack of visualfeatures on the XT space, which is caused by reducing the image dimensions fromXY to X, the proposed methods extract multiple layers of XT images from thevideos, and assemble them into a new2.5D data space. The following visual featureextraction methods are all based on this2.5D data space. The main works of thisthesis are listed as follows:1. Based on object motion feature extraction on the proposed data space, wepresent a spatial-temporal slice part model based object detection method. Thetraditional part based object detection methods need to pre-define a fixed number ofmultiple part models to represent object. Part detectors are built and trainedseparately, which induces hight computational complexity. To solve these problems,we propose a flexible and non-semantic part definition method. Every movingspatial temporal slice sub-region is viewed as a candidate object part. An unifiedpart detector is proposed to detect all the candidate object parts. Then, a clusteringalgorithm is proposed to combine these parts into individual objects based on theirconsistent motion patterns. The proposed method has low computational complexityand satisfies real-time object detection requirement. Benefiting from the flexible andnon-semantic part definition method, the proposed method can detect objects underpart missing or crowded environments. Furthermore, by introducing the multiple spatial temporal slices combination strategy, the method has the abilities ofdetecting varying number of multiple objects from any region of the video volume,and handling object enter and exit without additional detection mechanism.2. Based on object trajectory feature extraction on the proposed data space, aspatial-temporal slice trajectory detection based object tracking method is presented.Track shifting is a common problem in object tracking, due to similar backgroundscenes and object missing. In lack of long time object motion features, the algorithmhas no mechanism to revise these tracking errors. To solve these problems, wepropose a spatio-temporal slice complex trajectory detection algorithm to trackobject motions. Hough transform and subtraction clustering methods are employedto deal with repeat positioning, vertically moving and complexly moving problemsin the trajectory analysis process, which extends the traditional trajectory detectionmethod to handle complex trajectory patterns. The proposed approach fulfills objecttracking with complex motion patterns, including direction changing, velocitychanging, vertically moving and crossing. The approach is robust to short-timeocclusion, object pose variation and scale variation. By using the Hough trajectoryassociation strategy, the approach solves the track shifting problems caused byobject missing and tracking errors.3. Based on the proposed object detection method, a spatial-temporal slicecombination based object contour tracking method is presented. To solve complexobject representation and poor multiple object tracking ability problems in objectregion tracking, we propose a model-free tracking approach using a combination ofmultiple spatial-temporal slices. Object tracking problems in the3D video space aretransformed into object slice tracking problems in a new2.5D space, which iscombined by multiple XT spaces. A motion and visual feature based multiple framesmultiple slices corresponding approach is presented to track object slices in the newspace. The approach has the ability to track multiple objects, and provide a fineobject contour result. The proposed object tracking method deals with the problemsof occlusion, scale variation and multiple object tracking, satisfies real-timemultiple object tracking requirements. Besides, the method solves the track shiftingproblems caused by short-time or long-time occlusion, and is robust to object partmissing under environmental disturbance.