Research On Video Moving Objects Nonparametric Detection

Video moving objects detection is a basic research content in computer vision, the detectionresult of which plays an important role in objects recognition, image classification and so on. Thepopular moving objects parametric detection methods limit their adaptability to the complex anddiversified natural environment because it should assume background model. Compared withparametric detection method, the nonparametric detection method has better flexibility andadaptability to the environment. So video moving objects nonparametric detection under complexbackground is selected to research in this paper, and for goodness of fit testing and image texture,three novel nonparametric detection algorithms are proposed.Video moving objects detection using goodness of fit testing is formulated as a goodness of fittesting problem. Assuming background follows Gaussian mixture model, an empirical distributionfunction is introduced to model sample data statistically. The moving objects can be detectedaccording to the maximum distance of the two models. Performance of this algorithm usesKolmogorov-Smirnov（KS）test, which can judge whether the observed pixel is a background pixelaccording to the maximum distance between its empirical distribution function and backgrounddistribution function.Seeing that the minimum weighted KS estimates have good mathematical properties such asconsistency and robustness, video moving objects detection uses minimum weighted KS test tocalculate background model parameters, in order to eliminate the time limit of backgroundstatistical model and model deviation caused by small probability events happen and enhance therobustness of background model. The implementation of this algorithm exploits Bell-Curve Basedevolutionary optimization (BCB) algorithm to obtain the minimum weighted KS estimates ofbackground model parameters. This algorithm can improve the background model’s adaptability tothe natural environment because the acquisition of BCB is that the global optimum in thebackground model parameters space.Steering kernel regression-based video moving objects detection with the learned dictionariesapplies steering kernel regression function to represent background texture. It overcomes thedifficulty that flat region can’t be effectively described by traditional objects detection textureoperator. Background image is divided into several structurally similar clusters by use of K-meansclustering algorithm, and further Principal Component Analysis（PCA）is used to extract thecorresponding background texture dictionary. The learned background texture dictionaries are usedto reconstruct the observed image background texture, which can avoid environmental noise interference and enhance the flexibility and robustness of the background model. A detectionscheme uses two weak detectors to compose in series a strong detector, in order to detect movingobjects from cluttered background effectively. First moving objects in the observed image areextracted roughly by the weak detector based on pixel gray values. Then accurate moving objectscan be detected by the weak detector based on the learned background texture dictionaries tocompare actual texture and reconstruction texture.With Water Surface and Waving Trees video sequences, experimental results verify theeffectiveness of the proposed algorithms. Compared with the state-of-the-art detection methods, theproposed algorithms have the remarkable performance. Meanwhile, it is seen that all the threeproposed detection algorithms can detect the moving objects from the complex backgroundaccurately and efficiently, but their detection precision and performance improve gradually.