Study Of Visual Tracking Algorithm Based On CUDA

Visual tracking is an important issue of computer vision, and is widely used in many fields such as pattern recognition, machine learning, image processing and automatic control. There are two parts in visual tracking:object detection and object tracking. Object detection deals with detecting instances of semantic objects in videos. Object tracking deals with estimating the object motion in videos. There are many interesting applications of visual tracking in surveillance, medical diagnosis, intelligent transportation, perceptual user interfaces, augmented reality and driver assistance.Tracking learning detection (TLD) is an on-line visual tracking algorithm proposed by Zdenek Kalal at University of Surrey. It is an algorithm suitable for long-term tracking with good robustness and high accuracy properties. CUDA is a parallel computing architecture and programming model invented by NVIDIA.In this paper, the TLD algorithm is improved, and a fast algorithm based on CUDA is proposed, named CUDA-TLD. The CUDA-TLD algorithm has higher accuracy and better real-time performance. The main results are as follows:The improvement of TLD.First, the number of patches in the original algorithm becomes more and more during the tracking, causing bad real-time performance. The updating module of object model is improved by giving a threshold for the number of the patches in the object model. When the number of patches in the object model reaches the threshold, the patch with the weakest identification capability is replaced to maintain the number of the patches. As a result, the real-time performance is improved. Second, the detection module is improved. A detection algorithm based on outlier analysis is proposed, in which the outlier patches are classified as background patches. By integrating the algorithm into the detector of TLD, the detection rate is improved.The parallel implementation of TLD based on CUDA. The execution times of all the modules of TLD are compared, and then the detection module which is found to be the most time-consuming module is parallelized using CUDA, finally a fast algorithm based on CUDA is proposed, named CUDA-TLD. The experimental results show that CUDA-TLD runs faster than TLD for all the videos in the data set while promising the detection rate at the same time, and the speedup is between2.02and2.59. In particular, for the VGA videos with resolution640x480, CUDA-TLD can run at about18.7FPS while TLD runs at about8.4FPS, and thus satisfy requirements of the real-time performance.