Research On Visual Tracking Algorithm Based On Embedded System

Visual tracking,which determines the pose of object of interest through the techniques of image processing and filtering estimation in the image sequences,is one of active issues in the field of computer vision,and can be widely applied in military and civil applications.Currently the realization of robust visual tracking in complex scenarios is still a challenging problem,and the difficulties to be solved include target deformation,rapid movement,partial and full occlusion,and so on.The visual tracking algorithm based on correlation filtering has the advantages of good real-time performance and high tracking accuracy,and can be easily implemented in embedded system.In order to solve multi-feature fusion problem in correlation-filtering-based tracking and meet real-time requirement in intelligent surveillance application,the research on visual tracking algorithm based on embedded system has been done in this thesis,and the main research contents are listed as follows:Firstly,a robust correlation-filtering-based visual tracking algorithm based on multi-feature hierarchical fusion is proposed to improve the robustness of target tracking.Three features,including Histogram of Oriented Gradient(HOG),Color Name(CN),and color histogram,are extracted from the region of target and its surroundings to depict the appearances of the target and background when multi-channel correlation filtering algorithm is used to track the target.Adaptive weighted fusion strategy is firstly used to combine the response maps of HOG feature and CN feature at the first fusion layer,where fusion weights are computed by calculating Smooth Constraint and Peak to Sidelobe Ratio(PSR)of the feature response maps.Fixed coefficient fusion strategy is then adopted to combine the feature response maps when the first-layer fusion result and the feature response map based on the color histogram are combined at the second fusion layer.Finally,the position of the target is estimated based on the final response map,and the model update procedure is performed to adapt to appearance changes after scale estimation algorithm is used to obtain more accurate bounding box of target.Experimental results demonstrate that,the robustness of the proposed multi-feature hierarchical fusion tracking algorithm is superior to other algorithms under the premise of ensuring the tracking accuracy.Secondly,system design and algorithm transplantation of one embedded visual tracking system based on DSP+FPGA architecture are realized to meet real-time requirement in visual surveillance applications.The procedures of video capture and image preprocessing,including auto-exposure,color interpolation,median filtering,and white balance,are first carried out using FPGA as main controller.DSP processing results are exported into host computer for display through PCIE interface.High-quality images are obtained from FPGA to DSP through EMIFA interface,and the tracking algorithms based on meanshift and multi-channel correlation filtering are implemented separately in DSP.The tests for embedded system demonstrate that,it can obtain high-quality images,the target can be tracked stably and in time,and the system can meet real-time requirement.Finally,the main work and further research of the thesis is summarized.