Multi-appearance Models Adaptive Weighted Particle Filter Object Tracking Algorithm Research

Visual object tracking is an important task in the field of computer vision, it is designed to determine the position of the target object in the continuously changing dynamic scene. Object tracking technology has important research value and practical value, which is widely used in robotics, medical diagnostics, visual surveillance, military navigation, intelligent transportation and so on, involving multiple disciplines such as digital image processing, pattern recognition, machine learning and artificial intelligence, to name a few. In recent years, with the popularity of computer vision applications, the study of object tracking technology is unceasingly thorough, and has obtained certain achievement. The appearance change caused by numerous factors such as illumination, pose angle, occlusion, camera motion and so on, makes the current target tracking algorithm in terms of accuracy and robustness still face enormous challenges. Low-dimensional feature produced by partial least squares method can effectively describe the appearance variations of the target, at the same time, combine with the characteristics of the sample category, commonly used in the establishment of the target appearance model. But such tracking algorithm ignores the differences between the different dimensional characteristics, as well as the degree of association between appearance model and object, when the appearance of the target has complex changes, it is difficult to accurately analyze the error between target and sample, which reduces the tracking accuracy.In this paper, we carry out targeted research and analysis on the problems of current object tracking algorithm. Then, we propose an adaptive weighted object tracking algorithm based on multi-appearance models(AWMA), which accomplish the error analysis between object and sample by integrating of multiple appearance models. Specific work is as follows:(1)This paper presents a method based on feature weighting in a single appearance model, according to the importance of different dimensional features in appearance model, calculating the proportion of each feature in appearance model adaptively. This method can highlight the performance of the sample on different characteristics, more accurately describes the differences between sample and appearance model;(2)This paper presents a weighted method based on multi-appearance models, according to the degree of association between different appearance models and target, adaptively allocates the weight of each appearance model. This method accomplishes the error analysis between object and sample by integrating of multiple appearance models, to be able to reduce the tracking results affected by single appearance model.(3)In this paper, the appearance model update strategy adopts the combination of static and dynamic way, which keeps the appearance of initial target unchanged, and updates the appearance model based on the tracking results of subsequent frames, to make the tracking more adapt to changes in the real scene.In this paper, we compare AWMA and the current several tracking algorithms, and conduct experiment analysis through multiple sets of video data, to verify the robustness and effectiveness of tracking algorithm proposed in this paper.