Research On Small Object Detection And Tracking Algorithm Based On Surveillance Video

With the rapid development of artificial intelligence and machine learning,the widespread popularity of camera terminals,related research and applications based on computer vision technology have received unprecedented attention.Object detection and tracking is a basic and very important research field of computer vision,which plays a vital role in intermediate and advanced semantic analysis such as behavior recognition,event detection and behavior prediction.So far,the object detection and tracking technology has developed rapidly and has been widely used.It has achieved satisfactory results in terms of accuracy and real-time performance.However,in some application fields such as aviation and maritime affairs,when the object is far away from the imaging device or the object itself has a small scale,there are still a lot of problems to be solved for the detection and tracking of these small-sized objects,such as low resolution,background clutter,occlusion,tracking loss,low recall rate,etc.Aiming at the above problems,this thesis studies small object detection and tracking algorithms based on surveillance video,and proposes corresponding methods to solve the problems of low object recall in small object detection and tracking,so as to improve the effectiveness of video surveillance,the main contributions of this thesis are as follows.1.Propose algorithm BFSSD(icubic Interpolation and Feature Fusion based Single Shot Multi Box Detector Object Detection Algorithm).Taking the SSD network as the basic framework,the corresponding feature layers are up-sampled using the bicubic interpolation method;Then,the corresponding feature layers are fused;Finally,the validity of the proposed algorithm is verified.The algorithm is oriented to small objects,focusing on the context information fusion of feature layers,and using the key operation of feature fusion-bicubic interpolation to complete the work of dimensional alignment.In order to meet the detection needs of small and medium objects in the object detection,the SSD algorithm is optimized.The optimized algorithm fuses the underlying location information with the higher-level semantic information.Experimental results show that the algorithm improves the detection rate of small objects.The average detection accuracy of the PASCAL VOC2007 dataset is 79.3%,which is higher than the detection accuracy of SSD and DSSD(Deconvolutional Single Shot Detector).2.Propose algorithm FSCF(Feature Fusion and Scale Adaptation based Correlation Filter Tracking Algorithm).In order to solve the problem of insufficient object representation in the correlation filter,a method of the feature expression by fusing color attributes is proposed;Aiming at the problem of fixed template size in correlation filters,an effective scale adaptive algorithm is proposed.Tracking needs based on small video objects,the correlation filtering algorithm is optimized to enhance the feature expression of small objects.At the same time,small objects are detected at positions and scales.Experimental results show that the algorithm improves the tracking efficiency of small objects.Experiments on the VOT 2014 dataset,the proposed method can track about 78% of small video targets.The 50 sequences on the benchmark dataset OTB-50 are better than the CN,CSK,SAMF,and KCF trackers.