Tracking In Traffic Environment

Multiple target tracking is one of the most important module in intelligent transportation system, it is the key for intelligent traffic monitoring and other real-time visual applications. Traditional tracing methods usually are limited to a specific scene but difficult to deal with complicated environment changes, and the balance of the real-time and accuracy is also a problem.For meeting the demands of above aspects, in this paper, we take traffic environment video image sequence as the object, vehicle detection and tracking in video for content, with the purpose of a real-time and accurate target tracking method based on detect in the traffic environment.There are two main difficulties in varying number of targets tracking problems: First, observation model and targets distribution have pretty good chance to be nonlinear and non-gaussian; Second, it is easy to lead to some complicated situations when the number of tracking targets is changing in a scenario, such as overlap, shade, fuzzy, etc. In order to overcome these problems, this paper introduces a learning, detecting and tracking visual system for the interesting targets. It combines the advantages of two very successful algorithm: mixture particle filter and AdaBoost. The key points of the mixture particle filter design are suggested that the selection of proposal distribution and the solution when targets in and out. The learned AdaBoost vehicle detectors can quickly detect the vehicle targets that pass into the scene, and filtering processing allows us to continue the tracking of each individual player. Combining AdaBoost and mixture particle filter, and the auxiliary vehicles target tracking prediction for the occlusion issue, we get a powerful and automatic multiple target tracking system in all directions. This system does well in tracking vehicles in the traffic environment.The main work is as follows:(1) To avoid adhesion between the block mass and motionless targets(no foreground), in this paper, we use AdaBoost as the basic method of vehicle detecting, rather than the traditional target detection based on foreground segmentation. After learning and training the vehicle characteristics of complete directions,we construct a adaptive vehicle detection system.(2) On the basis of vehicle detection system, BPF framework is applied to the vehicle tracking, and according to the characteristics of vehicle in traffic environment, BPF algorithm is corresponding modified, making it more suitable for vehicle tracking.(3) on the basis of deep understanding of BPF algorithm, this article proposes a IBPF algorithm, which can solve several problems that BPF cannot, as the improved BPF algorithm, and there is corresponding solution in the situations of similar target or overlapping with each other.(4) The time of BPF algorithm performance is under the influence of target number and size, this paper we transform BPF algorithm to multithreaded IBPF algorithm in project implementation stages, making it not slow down because of the targets quantity rising.By aspects of theory and experiments, the pros and cons of multiple target tracking scheme are verified for kinds of the traffic environments, and based on this we put forward our own solutions. Hope that these designs and researches make their contribution for vehicles detecting and tracking technology in the traffic scene, and do my bit in the research of intelligent transportation field.