Research On Moving Object Tracking Algorithm Based On TLD Framework

As an important technology in the field of computer vision,object tracking provides the foundation for many advanced applications,such as gesture recognition,visual navigation,intelligent monitoring and so on.However,the performance of tracking is easily influenced by the changes of background and the pose variation of targets.The tracking framework Trackinglearning-detection(TLD)proposes a useful measure to solve the long-term tracking problem by decomposes the tracking problem into three tasks: tracking between frames,online learning and detecting in frames.Focusing on long-term robust tracking,the research on object tracking algorithm based on TLD framework and kernel correlation filtering tracking algorithm is launched.The main contents are summarized as follows:1.Multi-feature and high-confidence updating in kernel correlation filtering tracking are studied.Two specific tracking algorithms with various features and different combining methods are proposed.Firstly,by combining histogram of oriented gradient(HOG)and color name(CN)features at the feature layer,a high-confidence updated kernel correlation filtering tracking algorithm with fusion of HOG and CN features(HHC-KCF)is proposed.Then,the confidence of the tracking result is calculated based on the peak value and peak-to-sidelobe ratio of the response map.According to the confidence,the tracker is updated in the case of high confidence to prevent the tracker from corruption.The second tracking algorithm is a highconfidence updated kernel correlation filtering tracking algorithm with traditional and CNN features adaptable fused(HATC-KCF).Based on HHC-KCF,CNN feature is extracted to train a kernel correlation filter and then adaptively fused with the traditional at the response layer.On the basis of the fused response map,the position of the object is obtained.Additionly,whether to update the correlation filters is decided by the confidence of the fused response map.Finally,the performances of the proposed trackers are evaluated by multiple comparative experiments.2.An object tracking algorithm based on improved TLD is presented.First of all,the builtin tracking module in TLD is replaced by a multi-feature and high-confidence updated kernel correlation filtering tracking algorithm.Besides,the failure detection in tracking module is realized by the confidence of response and low-confidence stands for a tracking failure.Secondly,the searching strategy of detection module is optimized to reduce calculation.The area to detect is more specific and narrowed based on the estimated location of the object from the tracking module.Detection in full frame is not necessary until the tracking module fails to track.Thirdly,to avoid sample redundancy and pollutions,three sparse deletion rules for samples are put forward to remove the patches with weak representation ability from candidate samples and delete the redundant samples in object model.Experiments are conducted on the object tracking benchmark(OTB).Results show that the tracking accuracy and speed of the improved TLD are much better than before.What's more,the improved TLD exhibites excellent performance when comparing to other state-of-art tracking algorithms.3.A software of single target tracking system is designed and implemented.Primarily,requirement analysis is carried out to confirm the necessary functions.Then,the system is divided into six function modules: video sequence obtaining,target initialization,tracking process control,tracking result display,results saving and results anlysis.After that,the tracking system is developed by Python and QT.Then,the object tracking algorithms proposed in this paper and the classic tracking algorithms are integrated into this system.Finally,the robustness of the improved TLD when tracking in actual scenarios is verified through the system.