Research On Moving Targets Detection And Tracking In Complex Video Scenes

In the Intelligent Video Analysis System,the detection and tracking of moving targets is the key component of the process.The processing performance is directly related to the results of advanced semantic processing such as subsequent activity recognition and behavior understanding and so on.Because of the complexity of the video surveillance scenes and the multiplicity and the concurrency of interference factors,it is of great significance to study high-precision,low-time-consuming and strong robust moving target detection and tracking algorithms.Based on machine learning and the computer vision theory,the dissertation focuses on the detection and tracking of moving targets in complex scenes.The main work of the dissertation can be summarized as follows:(1)The illumination variation and moving shadow in the complex video surveillance scenes will produce a large number of false positive pixels in the moving targets detection result.Addressing this issue,the Local Binary Pattern is studied deeply and an improved local texture feature descriptor is proposed.Compared to traditional LBP,the improved feature descriptor has better noise robustness with gray scale invariance.Then,the texture feature is combined with the color feature to perform kernel density estimation and the probability model of the background is established.In order to improve the detection accuracy of the algorithm in the multi-modal background scenes,the neighborhood correlation of the pixels is used to effectively suppress the false foreground in moving targets detection.Experiments in standard test sets show that the method can improve the overall performance by 17.5% in slow illumination variation and motion shadows scenes compared with algorithms based on texture histograms.Comparing the other outstanding methods,the average detection performance is increased by 0.12%,while the processing speed is increased by 50%.(2)The traditional color space is easily distorted by the environment,the dissertation applies Color Name in moving targets detection and proposes a background modeling method based on regional Color Name spatio histogram.At first,the RGB color space is mapped to the robust Color Name space with reduced dimension.Then,the spatial structure of the pixel is introduced in the histogram and spatial histogram based on Color Name is established in the local area.And then,the background model is constructed by several spatial histograms with different weights.In the matching of the models,the similarity of the feature and the similarity of the spatial structure are both taken into consideration in order to improve the accuracy of the detection.The learning rate is used to control the update of each model histogram and its weight in detection.The parameters involved in the algorithm are discussed in detail.At last the feasible values of parameters are given.Experiment results show that the method has improved the overall detection performance by nearly 13% compared with other algorithms in dynamic background scenes.At the same time,the average overall performance of this method in 6 types of scenes is also optimal,reaching 75.18%.(3)Statistical background modeling algorithms treat each frame and each pixel as independent and unrelated individual so that inherent essence of high-dimensional data has not been explored.To solve the problem,the dissertation performs research in the moving targets detection algorithm based on Robust Principal Component Analysis(RPCA)and a detection algorithm based on adaptive low-rank and sparse decomposition is proposed.Firstly,the augmented matrix is constructed by the background model and the frame vectors to be solved.Then the dimension-reduced augmented matrix is decomposed by robust principal component analysis.The separated low rank part and sparse noise correspond to background and moving foreground respectively.At last,the background model is then updated with the currently obtained background vectors by the Incremental Singular Value Decomposition method.Experiments show that the proposed method can improve the robustness of model by introducing the adaptive process in the traditional principal component analysis.In three complex scenes with illumination changes,multi-modal backgrounds and bad weather,the overall detection performance F1 of this method is increased by 6%,8% and %7 respectively.Adaptive method effectively reduces the delay of the algorithm by loading video frames in batches and the time efficiency of detection in complex scenes is improved by 19.3%.(4)Aiming at the problem that appearance model of Particle Filter tracking algorithm is susceptible to environmental interference,particle impoverishment and particle degradation in particle importance resample,the dissertation constructs a particle filter tracking algorithm based on multiple cues and features for single target.A robust fuzzy statistical texture feature is obtained by fuzzy clustering to traditional Gray-Level Run-Length Matrix(GLRLM).It is combined with Histon histogram which reflects pixel correlation to construct appearance model of target.For the problem that the particle filter algorithm is easy to degenerate and deplete,the clustering algorithm is used to obtain an adaptive number of high-weight particles.These particles are also used to approximate more accurate target state estimation.In the importance resampling step,this part of the high-weight particles is retained and the remaining particles are resampled around the estimated target state.The mechanism overcomes particles degradation and maintains particle diversity effectively.Compared with other particle filter tracking algorithms with excellent performance,the average tracking success rate and the average tracking accuracy are improved by 4.4% and 2.8% respectively.Experiments also show that the proposed algorithm achieves optimal tracking accuracy and tracking success rate graphs in three scenes such as target deformation,illumination variation and background disturbance compared with several popular and superior tracking algorithms.