Study On Object Tracking By Detection

Letting the computer simulate human vision has always been an important branch of artificial intelligence, the increasingly explosive growth of video information generates strong demands for automatic analysis of the video. Among the automatic pipeline of video analysis, object tracking is a key technology, widely used in various fields of civil and military project, has a great applicational value, and becomes one of the cutting-edge research concern in computer vision. But object tracking is a challenging problem, with many difficulties such as the light changes, posture shape changes, background clutter, half or fully occlusion, blur and noise, these difficulties render tracking result in the drift phenomenon. The progress of object detection in the recent years makes tracking by detection technology an promising alternative for solving these difficulties. Tracking by detection trains an appearance model to capture the object appearance changes using machine learning methods. Tracking by detection generally contains three components:appearance model, motion model and search strategy, this paper focuses on appearance model, which is categorized to region-based feature method and keypoint-based feature method. Region-based feature focuses on statistical information of region, while keypoint-based feature focuses on local keypoint information. The different feature makes the appearance model different. Both kind of method still exists some problems need to be further explored. First, how to effectively extract feature for object representation, for object may be any kind and any form, the effective feature extraction is a difficult task; Second, the object appearance changes needs to be captured by an effective appearance model; Finally, how to reduce the probability of wrong update of appearance model in tracking process. This paper conducts several researches for these problems in tracking by detection:1. In order to better characterize the object and capture the changes in appearance, the paper proposes a tracking by detection method, which uses random fern classifier with compressed sensing based features as appearance model to represent objects. Compressive features retain most of the information and greatly reduce the number of features. The random fern classifier effectively combines the random characteristics of compressive feature, and efficiently learns object appearance changes. The final tracking results are based on optical flow tracker, the online random fern appearance model and the object templates. The experimental results show that these components compensate information for each other and the final results is better than used solely.2. This paper proposed a method of tracking by detection based on the key points features. The object model is defined as a set of weighted keypoint clusters, each keypoint cluster is clustering of similar keypoints. Boosting algorithm combines these weighted keypoint clusters as the object appearance model, and continuously updates it in the process of tracking so that effectively tracking the object appearance changes. The experimental results show that the method can provide stable tracking results for complex surface object such as human face and run in real time for taking advantage of binary keypoint description.3. Tracking by detection methods often suffer from wrong updates which could lead to drift or even failure, this paper proposes a method based on global optimization to optimize the tracking results. Drift is assumed as a time accumulation problem. Samples are collected using a weight descending strategy, an offline classifiers is trained and used to optimize the tracking results iteratively. The experimental results show that the final tracking results are more stable and precise than the original tracking results. This global offline classifier is robust to the fully occlusion and re-entering the scene, and overcomes the forgotten phenomenon in online learning. Unlike some other global optimization offline method, this method does not require any human interaction or additional labeling, can be used as a general-purpose optimization method to optimized the results of the detection by tracking methods, which could be applied in a post hoc analysis occasions.