|
Moving target detection and tracking is an important computer vision research, which combines computer image processing, artificial intelligence and pattern recognition, and many other fields of knowledge, and in the virtual reality, traffic monitoring, biomedical, intelligent human-computer interaction and other fields has a broad application prospects. Target detection based on the conventional process, the model, the model parameters and the model updates the sensitivity of the scene change, affecting the detected speed;, and moving targets affected by the light casts shadows produced will affect the detection of accuracy. In complex scenarios, because changes in illumination, moving target and background as well as moving targets high similarity crossover sport, etc., making the moving target tracking process will be tracked to be lost. To solve these problems, this paper complex scene moving target detection and tracking problem is studied, and its main contents and results are as follows: (a) Gaussian mixture model proposed self-learning algorithm through the traditional EM (Expectation Maximization ) algorithm, based on the learning rate factor is derived and forgetting factor recursive expression, making the parameter update is more accurate, faster convergence. For light cast shadows generated in the process of moving shadow detection has also been detected, the paper HSV (Hue, Saturation, Value) color model to eliminate shadows, making the complex context of the foreground object detection more accurate. Experimental results show that the conventional Gaussian mixture model needs to a good 40 or so to detect moving targets, and Gaussian mixture model self-learning algorithm in the first 22 or so can be very good to detect moving targets, and through the shadows eliminated, the foreground object is also more accurate. (2) made of interactive MCMC (Markov Chain Monte Carlo) particle filter, to solve the problem of lack of particles, by introducing the particles to reduce interaction between the particles of the history information on the link between the state, the number of particles not only solve the problem of the degradation, but also accelerate the convergence rate and improve the performance of the algorithm. Body-movement detection using the result, select the target area containing two diagonal vertices of the boundary as the tracking feature points, with the particle filter to predict the interaction MCMC and track the position and speed of the feature point, to obtain the trajectory of the feature point, the feature point selection the three dimensional position and velocity as state variables, thus avoiding direct linear function of the nonlinear tracking errors caused. Experimental results show that in a complex environment for target tracking, even when there is light, high similarity target and background circumstances, it does not appear the phenomenon of lost track targets, while also able to reliably predict and track human movement in 3D space trajectory. (3) proposes a fuzzy data association and particle filter combination of methods, and for multi-target tracking process, the method by resampling particle filter after adding improved fuzzy membership function will be to get the best data association membership degree as the particle weights, so as to effectively avoid the interference of noise data, and through the performance of the algorithm analysis, the cross-track the target will not be lost when the tracking target. Experimental results show that in a complex environment, movement occurs when the body can also be a good cross when tracking the human body, and will not appear the phenomenon of lost track, thus increasing the target tracking robustness and accuracy.
|