Research On Improved Tracking Algorithm Based On TLD Algorithm

Target tracking remains an important research topic in the computer science field with ubiquitous practical support in human-computer interaction,intelligent security,visual effects,autopilot,weapon guidance,and other applications.It has also extensive adoption in intelligent transportation,video intelligence analysis,pedestrian behavior analysis and many other fields,paving its way to a wider range of applications.Yet there are many challenging and difficult circumstances in target tracking: a large amount of data and the complexity in calculations,partly exhibiting the intrinsic characteristics of the video sequence itself.So the real-time of target tracking was somehow impossible;also there are contributing factors such as the complex video content scenes,smog,intense illumination changes.Many disturbance factors such as fast motion and motion blur exist,making applicability the second largest problem for target tracking.Last but not least,unpredictable interference may occur randomly like the occlusion,scale change,rotation,and sudden change in background.Therefore it's often difficult to balance tracking success rate and tracking speed.In this paper,we deeply studied several tracking algorithms with extensive comparison experiments through public data sets.Many advantages and disadvantages were revealed for the algorithms and strategies,and many valuable research results were obtained.Key points for the main work of the paper are as follows:1.Based on Retinex algorithm,combined with similar block and wavelet threshold filtering,an image defogging denoising algorithm for image fogging and noise problem is proposed.By decomposing the image,estimating the incident light component,solving the reflected light component,and achieving the image defogging effect,further dividing the image into blocks,and stacking blocks smaller than a given threshold into groups according to the Euclidean distance between the blocks,in the group The image is filtered by wavelet transform domain to remove noise,and then the image block is returned to the original image by weighted averaging to achieve the overall defogging and denoising of the image,which provides high-contrast and lownoise high-quality images for feature extraction of subsequent tracking algorithms.2.In-depth analysis of TLD tracking algorithm,proposed to use KCF algorithm to replace TLD algorithm tracker,improves the real-time and applicability of TLD algorithm.Further analysis of the KCF algorithm to deal with the target size changes,occlusion and similar object interference,using VGG-16 deep neural network different layers can extract features of different characteristics,the different layer features are combined for tracking,to solve the KCF algorithm target size The tracking failure caused by factors such as scale variation,occlusion,and similar object interference improves the robustness of the KCF algorithm.Improving the real-time performance of the tracking algorithm under the premise of ensuring accuracy.The equal interval frame update strategy was used to compare the tracking accuracy and speed of the tracker under different frame intervals.Based on the experimental data,the sparse tracking model constantly updates the strategy,reduces unnecessary calculations,results in further improvements of the real-time performance of the algorithm while keeping a certain tracking accuracy lower boundary.3.Combining the SSD network model and build a simple program for pedestrian tracking using the improved TLD tracking algorithm.We utilized the detection via the SSD network model,feeding forward the identification results generated by the network to our proposed tracking algorithm which would then work the tracking job out.