Study Of Pedestrian Tracking Method Combined Pedstrian Detector And Color Model

With the development of technology, the product of artificial intelligence is gradually getting into people’s lives, among which the pedestrian tracking technology has gradually attracted people’s attention and played an irreplaceable role in many fields. Nowadays, many cars have installed pedestrian stabilization warning system of which the pedestrian detection and tracking technology play a core role. Besides, pedestrian also play its role in security monitoring, behavior recognition and robot intelligence. Overall, pedestrian tracking algorithm is a technical issue related with many disciplines such as image processing, computer vision and pattern recognition. Although there are many preliminary research results about it, there still exists many theoretical and technical problems we need to solve. In order to conduct pedestrian tracking for a long period, there are a lot of problems needed to be solved. A relative good tracking algorithm must be able to handle problems such as scale, illumination changes, complex background and partial cover. Sometimes, target would also have some changes which would make the original mode lose efficacy, for example, deformation. In this case, the problem would be more complex. Also, tracking for a long period needs to solve the problem of losing target. Due to the fact that situations such as blocking and losing sight are inevitable, algorithm must have the ability to get the target back when it appears again. Of course, as a tracking algorithm, the real-time of it is also very important.This thesis aims to put forward a practical real-time pedestrian tacking algorithm which could relatively better solve the problems mentioned above. Aiming at this goal, a pedestrian tracking method combined pedestrian detector and color model is proposed in this thesis. The main research contents includes the following several aspects:1)Based on the target tracker with median optical flow method, an autoscope for tracking failure is added into the tracking process. The target tracker would track the target pedestrian and estimate his movement. In the process of tracking, the autoscope would estimate the reliability of the points having been tracked and then select those points with higher reliability to estimate the pedestrian’s location and scale of change. Compared with other tracking methods, median optical flow method not only could guarantee the accuracy but also have the advantage of running fast, which are important for the instantaneity of the algorithm.2)The pedestrian detection algorithm proposed in this thesis consists of two parts which respectively are the SVM pedestrian detector based on the features of HOG and the pedestrian detector based on the deep model and trained by the deep learning method. The former one has a faster speed but has problems such as low accuracy and poor stability, which would bring problems in the algorithm. Also, the latter one has a higher accuracy but the testing process takes longer time for there are a lot of convolution computation in the calculation process, which would seriously influence the efficiency of the algorithm. Based on the respective characteristics of these two algorithms, we combine them together by using a strategy to make the algorithm could has better performance both in speed and accuracy.3)This thesis proposes a pedestrian color model with recognition capacity. And also, considering that the target being tracked would change in its posture, complex background and covering during the tracking process, the updating strategy of the model has also been proposed to enable the algorithm to constantly adapt to the changes of the target. In the tracking process, the algorithm would combine the target tracker, pedestrian detector and the color model with recognition capacity together, letting them learn from each other and promote each other, so as to achieve the purpose of tracking pedestrians stably. Experimental results show that the algorithm proposed in this thesis has good stability and robustness in handling illumination variation, complex background, target deformation, covering and target missing.