Object Detection And Tracking For Intelligent Visual Surveillance System

Intelligent visual surveillance is an emerging technology focused all around the world. It can be applied into great deal of aspects including public safety protection, medical care, traffic management, customer service and many other fields. So it is urgently needed by cites with fast growing population and explosively expanding scale. It involves many subjects like computer vision, image/video processing, pattern recognition and artificial intelligence. The tasks intelligent visual surveillance has to complete include:moving object detection, object classification, object tracking, object behavior understanding and identity recognition etc. Two of these key issues are focused in this paper. They are the issues of object detection and object tracking. The main research work includes:1. In the aspect of object detection, an improved texture-based method using local binary patterns (LBP) for background subtraction is proposed. The original texture-based method itself is a successful solution to background subtraction especially for dynamic background scenes. However, there always exists relatively large number of false positives in the segmentation results near the edges of moving objects and it suffers from slow adaptation to the changing background. To solve the first problem, a spatially weighted LBP histogram is proposed to be the feature vector and a shadow removing method is introduced. When dealing with the second one, the uniform fixed learning rate in the original method is replaced by an adaptive learning rate calculated for each model LBP histogram and multiple frame-level models are maintained to handle sudden illumination changes. Experimental results show that the proposed method successfully solved the two problems in the original method without introducing any adverse effect.2. As for the topic of object tracking, a real-time robust visual tracking method which can be used in complex situations is presented in this paper. In this method, the color cue and the shape cue are integrated and utilized to represent the object tracked and further incorporated into the framework of particle filter (PF). In order to alleviate the heavy computational burden suffered by traditional PF, each particle of PF is moved by mean shift (MS) algorithm independently after the step of importance sampling. This increases the likelihood of each particle and makes them become more effective to approximate the posterior probability density function such that the total number of particles needed is significantly decreased. However, this modification has three drawbacks including:(1) The color feature on which MS depends may become unreliable so the particles may be moved to even worse positions; (2) MS itself needs relatively large amount of computation; (3) The diversity of particles is decreased by MS. In order to overcome the three problems, the maximal number of MS iterations is limited proportional to the reliability of the color cue computed by democratic strategy. Experimental results show that the proposed method outperforms the traditional PF-based methods with much smaller number of particles.