Research On Robust Appearence Modeling For Tracking Algorithm Of Visual Object In Complex Environment

Visual object tracking is an important research field in computer vision,which has a wide application for human-computer interaction,military defense,medical assistant diagnosis,autopilot system,and so on.Nowadays,with the development and application of the artificial intelligence,visual object tracking is not only fulfiled the basic tasks of feature extraction,detection and recognition,but also provides reliable input information for the higher-level vision analysis,such as target behavior analysis,trend prediction,semantic understanding,etc.However,the performance of the object tracking algorithms is usually affected by heavy occlusion,background clutter,motion blur,illumination,shape changes and other factors.Although a large number of excellent tracking algorithms researches has proposed in recent years,it is still an important and challenging task to study robust object tracking due to the influence of a complex environment.Aiming at the single object tracking scene of background environment interference and target apparent change,as well as the multiple object tracking scene with background clutter and similar target interference,this paper analyzes the shortcomings of the existing tracking technology in adapting to the above complex environment.Then further expands the research of multiple instance learning,multi-view feature fusion,Stacked Denoising Auto Encoder,Principle Component Pursuit,and other fields,which proposes a series of new optimization methods of appearance modeling in order to improve the accuracy and robustness of the tracking algorithm.The main contributions of this research are summarized as follows:(1)For multiple instance learning,the algorithm can not effectively distinguish the influence of different positive and negative instance sample on appearance modeling,which leads to discriminative ambiguity problems in background environment interference.A visual object tracking algorithm based on objectness measure with multiple instance learning is proposed.Based on multiple instance learning,the importance of each instance sample is evaluated by using the objective measurement criteria.And each positive instance is given the corresponding weight according to its objective measurement results,which is utilized for computing the final bag probability and improve the tracking accuracy.At the same time,in the phase of weak classifier selection,the method of maximizing the probability inner product of weak classifier and likelihood function is used to select weak classifiers from weak classifier pool,and then these weak classifiers are combined into a strong classifier.The optimization method reduces the computational complexity of the algorithm.By tracking video sequences in a different complex environment,the experimental results show that the tracking algorithm of objectness measure with multiple instance learning proposed in this paper is better than its comparison algorithm,and shows better tracking accuracy and robust performance.(2)In view of the fact that the single view feature description can not adapt to the background environment interference such as illumination and occlusion,which leads to the loss of some important information on the image during the feature extraction.This paper proposes a tracking algorithm based on multi-view feature fusion with online multiple instance learning.Firstly,both of the color histograms and Haar-like feature in the image are extracted to construct two types of weak classifiers,which describe the appearance model of the object jointly.Then the weak classifier is selected by the method of maximizing the log-likelihood function of the bag package,and the strong classifier is constructed according to the features.Finally,the object location is achieved by fusing the two type of strong classifiers based on confidence weighting.The experimental results show that the multi-view fusion tracking algorithm has a better center position error in the video sequence of a complex environments such as occlusion and illumination change,and shows a better tracking performance.(3)Because of some hand-craft feature descriptions are easy to be filtered along with appearance variations such as shape change and motion blur.So in this paper,a stacked denoising autoencoder network is proposed,which uses the multi-layer neural network to model object appearance features.The self-learning process of the SDAE neural network contains two stage: In the forward propagation stage,the input image is represented by the double-layer sparse dictionary coding method.After training layer by layer,the feature descriptors of different spaces of the image are obtained to express the structure pattern and internal information of the image more clearly.At the same time,in the next reverse fine-tuning stage,by using a genetic algorithm to update the network connection weight and bias terms,instead of the BP algorithm that may fall into local minimum due to the lack of backward gradient propagation.Finally,the tracking algorithm is well-established by the particle filter network.Experimental results show that the target tracking algorithm based on the SDAE neural network can extract all kinds of image features in the hidden layer more effectively,and adapt to the dynamic appearance change of the object.Compared with other tracking algorithms,the quantitative indicators such as average center pixel error and average overlap rate have better performance in the test image sequences.(4)In TBD based multi-target tracking,due to the detection response provided by DPM detectors in a complex environment such as background clutter and similar object interference has missed detection and false detection problem,so this paper proposes an online multiple object tracking algorithm based on the overcomplete detections set.Firstly,this algorithm uses the sparsity of foreground target and the low-rank of background information to segment the foreground and model the background based on the optimized PCP,which embeds the predicted position of the object as feedback information into the low-rank sparse computing to achieve more accurate performance.Secondly,the detection response provided by the DPM detector is corrected by foreground appearance segmentation results.At the same time,the algorithm further uses the prior template knowledge of the tracked object to match the detections again and output the final overcomplete detection response set,which helps to eliminate the problem of false and missed detection response in multiple object tracking.Finally,in the tracking algorithm,data association is established in two stages based on the tracklet confidence,the fragmented tracklet are associated step by step to improve the accuracy of multiple object data association.The experimental results show that the multiple object tracking algorithm based on overcomplete response set significantly reduces the number of FP and FN,and improves the Recall and Precision performance.