Research On Adaptive Online Multi-filter Tracking Algorithm

Correlation filter based object tracking has recently gained popularity due to continuous improvements in the tracking accuracy and robustness.However,these trackers are limited by the serious model drift problem in some complicated tracking scenes.And the model drift increases as more frames are processed,which restricts the ability of long term tracking in correlation filter trackers.Therefore,this paper studies the influence of difficult problems in complex scenes on the correlation tracker.The main research contents and achievements are as follows:(1)A scale adaptive multi-filter tracking algorithm against occlusion is designed.The algorithm divides the samples into different sets based on their similarities.An independent filter is trained for every set.In the tracking process,the matching filter is selected for the current target by calculating the similarity between the candidate region and the sample set.In this algorithm,multiple templates stored in multiple samples independently.This algorithm improves the shortcoming of single filter which stores limited target information and has poor adaptability to target deformation.And it can effectively adapt to the discontinuous appearance model of the target.At the same time,according to the target verification and occlusion detection strategy,the target state can be effectively evaluated.by calculating the peak signal to noise ratio of the output response and regional color histogram of the tracking results,the tracking failure or occlusion of the current frame can be judged.And the tracking model can be dynamically updated according to the reliability of the tracking results,so as to further improve the robustness of the tracking model.In addition,in order to adapt to the change of the target scale,first of all,estimating the trend of target change roughly by three scale factor.Then according to the change direction to establish the scale space and get the best scale of the object.Finally,the experimental analysis of the algorithm and comparison algorithm is carried out on the OTB data set,which is proved that the algorithm can achieve adaptive tracking of the target scale,and is more accurate and successful than the comparison algorithm in complex scenes such as the deformation,partial occlusion,illumination,and so on.(2)A decision model of target state triggering is designed.Usually the tracker will fail to track the target when it occurs heavily occlusion or out of view.Therefore,this paper designed a redetection module for the above situation,combined the detector with the above tracking anomaly detection mechanism,which realized accurate tracking when the target reappeared after complete occlusion.In order to improve the accuracy of the redetection model,this paper extracts the detection samples by segmenting the whole frame image.The variance classifier and the nearest neighbor classifier are used as cascade detectors to classify the detection samples.At the same time,the multi-filter template is used as the learning sample to assist the target repositioning.Then the transition between the detector and the tracker is triggered based on the output of the detector,which improve the robustness of the algorithm in complex scenes.Finally,by selecting the corresponding test sequence in the OTB data set to evaluate the algorithm in this paper,the average success rate of this algorithm is 0.6929,the average accuracy is 0.7408,and the frame rate is 27.91.The experimental results show that the redetection module can effectively capture the target in the case of severe occlusion or tracking anomaly.Compared with the comparison algorithm,the tracking accuracy and success rate of the algorithm in this paper are both optimal,and can basically achieve real-time,which proves the robust tracking performance of this algorithm.