Research On Target Tracking Based On Algorithm For Occlusion Detection

With the development of artificial intelligence in recent years,it has received a lot of attention for computer vision.In the application scenario of target tracking,there will appear various challenges,such as occlusion,deformation,plane rotation,illumination variation,background clutters and so on.Nowadays,there are many target tracking algorithms,although they have their own strengths,there are still areas for improvement.In many algorithms,during the tracking process,the DAT algorithm(Distractor-Aware Tracking)detects the interference area similar to the target and the surrounding area of the target through color characteristics,and then combines by their weights,This effectively reduces the "drift" phenomenon of traditional color features in the tracking process,and achieves the tracking of targets similar background.However,when the target is occluded,the tracking will fail due to the loss of target information.In view of the deficiencies of the DAT algorithm when the target is occluded,based on the original DAT algorithm framework,this paper proposes a target tracking algorithm based on occlusion detection——DDAT(Detection-DAT)algorithm.The DDAT algorithm introduces an occlusion detection mechanism,which first extracts the color features of the target,establishes a target similarity model,and uses the similarity model to calculate the similarity of the target between frames,then determine whether the target is blocked by the threshold of the similarity difference.When the target is occluded,the detection mechanism is activated,and extract the target similarity model that occludes the previous frame.In subsequent frames where occlusion begins,use the Naive Bayes Classifier to get a better candidate bound,and finally get the best target area in the candidate bound by the Nearest Neighbor Classifier.Extract the color features of the target in the final area to establish a similarity model,and compare it with the target similarity model in the previous frame,calculate the similarity of the two frame target model,and use the difference in similarity to determine whether the target area is correct.This article mainly carried out research work in the following aspects:1.Explained the research background and significance of computer vision,introduced the development process of target tracking technology and the current status and challenges of tracking algorithms,and combed the research methods of occlusion problems.2.Discussed some basic theoretical knowledge of target tracking,and briefly explained the tracking process of the DAT algorithm,pointed out the advantages and disadvantages of the DAT algorithm,and proposed improvements to the deficiency of the DAT algorithm in occlusion.3.Based on the DAT algorithm framework,a DDAT algorithm with occlusion detection is proposed.Define the similarity trend,and the start and end of target occlusion are judged by using the characteristics of phasing and gradation of the occlusion.Use the occlusion phase and gradual characteristics to judge the start and the end of the target occlusion,and use the Naive Bayes Classifier and the Nearest Neighbor Classifier to detect the position of the occluded target,and finally the similarity is used to discriminate the accuracy of detection to achieve effective tracking of short-term occlusion.4.The qualitative and quantitative experimental comparison and analysis were performed on the OTB(Object Tracking Benchmark)standard data video sequence set with occlusion properties.The experimental results show that the DDAT algorithm effectively handles the tracking failure caused by occlusion.