Video Moving Target Tracking Algorithm In Complex Background

The tracking of moving targets has always been an important research topic in the field of computer vision.In the actual application process,it wills not only face changes in the object pose,changes in illumination,mutual obstructions between target objects,rapid movement of the target,but also need to solve the accuracy,robustness and realtime performance of the tracking algorithm.This paper presents two algorithms for how to track a target object under complex tracking conditions:moving target matching tracking algorithm based on effective feature points and video target tracking algorithm based on the pseudo-siamese network.TLD(tracking-learning and detection)algorithm is a kind of algorithm that can track single target objects for a long time.It combines tracking and detection,and introduces an online learning mechanism.All of that allow it learns the characteristics of the target object quickly and tracks effectively with little prior knowledge.However,TLD algorithm also has some obvious deficiencies,such as: when the TLD algorithm detects feature points of the global image,it will introduce some negative targets due to the large detection area,leading to long time for detection.The tracking module of the TLD algorithm evenly selects points for tracking on the target,but these points cannot effectively represent the target object.On the basis of the TLD algorithm,we make some improvements and propose a moving target matching tracking algorithm based on effective feature points.In the TLD detection module our algorithm add a directional predictor that can predict the direction of the moving target object,which can reduce the detection range of the detector and improve the ability of the tracking algorithm to recognize similar target objects.At the same time,we improve the selection of effective feature points and the layout of the tracker in the tracking module.The experimental results show that this method improves the robustness and accuracy of the tracker,especially when the target object moves quickly;it can reduce the average center position error effectively.At present,there are many algorithms only use the detector to track the target object.However,due to factors such as excessive calculation amount and complex tracking environment,it is impossible to track the target object in real time.In order to enable the detector to meet the requirements of the actual tracking process,we propose a video target tracking algorithm based on the pseudo-siamese network based on the deep learning.In order to improve the tracking accuracy and robustness of the algorithm,we construct a pseudo-siamese network by learning a large number of tracking data set and construct a loss function to optimize the model parameters at the same time.In addition,the high-level features extracted from the network have strong stability to the target change and can be used in the inter-classification objects;The low-level features extracted from the network retain a large amount of detailed information of the target object,which can be used to classify objects within the region and achieve effective differentiation of similar target.In order to improve the tracking speed of the algorithm,we use offline data to train the network and do not require online fine-tuning.In addition,the use of the center of gravity-based detection strategy to find candidate areas of the target,makes the calculation amount of the algorithm is greatly reduces,and the target object can be tracked in real time.The traditional convolutional neural network needs to input the same size of the picture,usually the input image needs to be cropped or zoomed,but these operations will cause the image data to be lost or geometrically deformed.In order to enable the network to input any size image,we add the spatial pyramid pooling layer into the network.By qualitative and quantitative analysis of the algorithm and the current mainstream tracking algorithm,our algorithms have higher tracking precision,higher processing speed,stronger ability to identify similar target objects and can track the target object robustly under complex conditions such as rapid movements of the target object and deformation of the target object.