Research On Target Tracking Algorithm Based On TLD Framework

As the core content of computer vision,target tracking has attracted more and more scholars' attention and research.With the rise of artificial intelligence,it plays an increasingly important role in intelligent nursing,video surveillance and military fields.However,due to the complex environment in the real scene,there are still many problems in target tracking,such as occlusion,deformation,scale variation,background clutters,illumination variation,rotation,motion blur,fast motion,beyond vision,etc.Making tracking still difficult to implement in any realistic scene.This paper mainly analyzes and studies the TLD(Tracking-Learning-Detection)algorithm suitable for long-term tracking,and improves the illumination variation,occlusion and background clutters based on the framework,and proposes fDS-TLD algorithm.In TLD algorithm,the performance of the tracking module algorithm in the original algorithm is difficult to deal with illumination variation,occlusion and background clutters.So an improved TLD algorithm based on fast discriminant scale space tracking(fDSST)is proposed in this paper.A fast discriminant scale space algorithm with dual filters of position and scale is used to replace the optical flow method of the original tracking module,thereby improving the overall performance of the algorithm.And after the algorithm replacement,a new tracking failure detection mechanism is adopted to make the replacement algorithm more integrated with the TLD framework.Aiming at the problem that the occlusion target can still be tracked after the algorithm is improved,the detection module is mixed into the occlusion sample to cause the drift of the target.This paper uses the perceptual hash algorithm to design the occlusion discrimination mechanism.The mechanism judges the target sample of the learning module to be successfully tracked.Only when it is judged to be unoccluded,it is updated as a positive sample to the detection module,otherwise it is not updated,thus ensuring the reliability of the positive sample of the detection module and improving detector performance.In view of the problem that the detection module adopts the global scanning method during the detection,the detection module takes too much time.In this paper,the Kalman filter method is added before the detection module to predict the possible area of the target,and then within the prediction range.By performing detection,many meaningless detections for background image blocks can be reduced,thereby improving the detection speed.At the same time,in order to ensure the original detection performance of the detection module,the global scan is performed again when the overall algorithm tracking fails to avoid the inaccuracy of the Kalman filter.Since in most cases the algorithm can track the target,in general,this method speeds up the overall algorithm.Finally,in order to prove the effectiveness of the improved algorithm,this paper compares the original TLD algorithm,fDSST algorithm,KCF algorithm and LMCF algorithm in the OTB-100 data set.Experiments show that the proposed algorithm is 25.5%higher in tracking accuracy than the original TLD algorithm,and is 30.3%higher in tracking success rate,and it is even better than other algorithms.At the same time,the tracking speed has reached 25.63 fps,which meets the real-time requirements,which proves the robustness,accuracy and feasibility of the proposed algorithm.