A Study Of Moving Object Detection And Tracking Based On Learning Method

The detection and tracking of moving object based on computer vision are widely used in military and civilian scenarios,such as video surveillance,human-computer interaction,image interpretation,unmanned aerial vehicle landing and so on.Therefore,they are always research hotspot in the field of computer vision.In the practical application,the detection and tracking of moving objects are inseparable and need to be combined to complete the specific tasks.Although the related research progress is encouraging,their efficiency,accuracy and automation still have room for improvement.As to moving object detection,this paper proposes a moving detection network with optical flow aiding.The optical flow deep learning network is trained to calculate the target motion between two frames,and then adaptive threshold is automatically selected to segment moving objects,which generate series of samples for background subtraction network.Finally,a slight background subtraction network is trained.During the detection,we only use the background subtraction network and avoid the time-consuming light flow computation.This method eliminates the dependence of the existing background subtraction network on the hand-labeled target mask and improve the automation of moving object detection.Aiming at the moving object tracking,this paper proposes a tracker based on support vector regression and design its online learning algorithm.We also come up a tracking method by fusing structure support vector machine method and correlation filter method.We concatenate these two trackers and design an alternate learning strategy of updating two kind of popular trackers simultaneously to combine the advantages of the structure support vector machine method in re-detection of occluded target and correlation filter method in scale estimation,which improves the precision and robustness of tracking.