PTZ Camera Pedestrian Tracking Research Based On The Improved Algorithm Of TLD

Target tracking is an important branch of the computer science field.It is the necessary conditions to achieving artificial intelligence,which playing an important role in artificial intelligence's more advanced processing applications.Pedestrian target tracking is the key and important first step in applications such as video surveillance,vehicle-assisted driving and human behavior analysis.Based on the pan-tilt-zoom network camera's active target tracking method,the tracking system can simulate the function of the human's eyes,the camera can follow the target's movement to make the appropriate action,so that the target is always kept in the camera's vision.The application of active target tracking method improves the shortcomings of traditional camera surveillance field.The tracking-learning-detection algorithm is a kind of target tracking algorithm which is commonly used in the long-term tracking.Applying the tracking-learning-detection algorithm to the pan-tilt-zoom camera tracking,can make up the problem of pan-tilt-zoom camera lost the target in a long time.But tracking-learning-detection algorithm has the defect and shortcomings.When the pedestrian target is affected by light,occlusion and other factors,the tracking accuracy of the algorithm is reduced,the complexity of the algorithm is high,and the real-time performance is poor.In this paper,aiming at shortcomings of the tracking-learning-detection algorithm to improve the algorithm,the main works are as follows:(1)Set up the experimental platform,select the center offset distance,overlap degree and average frame rate as the experimental evaluation index,and select 6 data sets from the Visual Tracking Benchmark data sets as the experimental analysis data sets.Analysize the experimental results,and establish the improvement scheme according to the shortcomings of tracking-learning-detection algorithm.Respectively,improve the feature description algorithm and classifier of the detection module,and the median optical flow method of tracking module,improve the sliding window method.(2)Select five common feature description algorithms,apply the five kinds of features to tracking-learning-detection algorithm,and analyze the tracking-learning-detection algorithmbased on 5 features experimentally.Combine the histogram of oriented gridients feature with the support vector machine classifier to improve the 2bitBP feature and the integrated classifier in the tracking-learning-detection algorithm detection module.Analyze the improved tracking-learning-detection algorithm based on experiment of the standard data set.(3)Select five kinds of common target tracking algorithms,and analysize the six kinds of tracking algorithms based on the experiment of the Visual Tracking Benchmark dataset.Improve the tracking-learning-detection algorithm based on the kernelized correlation filters tracking algorithm,and analysize the improved tracking-learning-detection algorithm by experiment.(4)Establish pan-tilt-zoom camera control model and pan-tilt-zoom camera dynamic tracking control system.Improve the sliding window algorithm and applied to the tracking-learning-detection algorithm based on HOG-SVM detection and kernelized correlation filters tracking algorithm.Apply the improved tracking-learning-detection algorithm to the experiment on a pan-tilt-zoom camera tracking.The improved tracking-learning-detection algorithm preserves the characteristics of the standard tracking-learning-detection algorithm in a long time tracking.On the basis of the standard tracking-learning-detection algorithm,the tracking precision and real-time performance are improved to some extent.