Optimization And Realization Of Video Object Tracking Algorithm Based On Kernelized Correlation Filters

Video Object tracking is one of the popular fields of computer vision research, it is widely used in various fields such as human-computer interaction, video surveillance, intelligent transportation, military attack and medical diagnosis. Video object tracking algorithm is divided into tracking-based algorithm and detection-based algorithm. Tracking-based algorithm is to use the information in the previous frame to estimate the object position and state in the current frame, this methods accumulate error during run-time (drift) and typically fail if the object disappears from the camera view; detection-based algorithm is to determine the position and state information of the object by the scanning of the current image,the methods need to train classifier offline and prone to false detection and missed the case. In the process of tracking, there are a series of problems,such as illumination change, scale transformation, appearance deformation and object occlusion or disappear, whether tracking-based algorithm or detection-based algorithm is unable to complete long-term tracking tasks independently.In this paper, after a deep research on the related video object tracking algorithm, I propose a long-term video object tracking algorithm based on kernelized correlation filters -- K-TLD (Kernelized correlation filters based Tracking-Learning-Detection). Using the kernelized correlation filters as a tracker, which is tracking speed and easy to achieve real-time tracking. The detector can use known information to solve the problems such as the object disappearance that can not be handled by the kernelized correlation filters tracker, with a learning component to improve tracking accuracy.Experiments show that the proposed algorithm can resist drift and anti-occlusion remarkably.The limitations of computing resources of Mobile Embedded platform makes it difficult to meet the real-time requirement by transplanting algorithm directly. After analyzing the time-consuming of each module of the K-TLD, using the NEON's powerful parallel instruction set to optimize the feature extraction module to satisfy the real-time requirement of applications.