Research On Pedestrian Tracking Algorithm

Object tracking is to find out the target and keep tracking of the target in a given video. This paper has studied the tracking of pedestrians, which has the most practical significance and application value among all the object tracking. Usually, object tracking algorithm can be roughly divided into detection based tracking and prediction based tracking, but any algorithm has its own defects and shortcomings, it deos not exist a universal algorithm. For example, detection based algorithms are often vulnerable to the influences of occlusion and deformation. Prediction based tracking algorithms make prediction error between the tracking result and the real target in tracking process, and the error will accumulate over time, making the tracking effect worse with time. On the other hand, pedestrian features are mostly from the experts' carefully design, but these features belong to shallow features, their description ability is limited. According to the above two problems, this paper proposes the corresponding pedestrian tracking algorithm, the main research contents and innovative contributions are as follows:(1) Proposed An improved TLD algorithm for pedestrian tracking. TLD algorithm proposed to combine tracking, learning and detection organicly, make up the tracker's error through the detector, update the detector with both positive and negative training samples, which generated through the online learning module, and improve the accuracy of detection. It shows a good effect in long time single target tracking. The TLD algorithm detects the target through a fixed sequence of three classifiers in the detection module, which can not adapt to changes in the appearance of the target well, so that the accuracy of the detection would be affected. We propose to improve the detector module by using the online AdaBoost algorithm to dynamically select the best features to construct a strong classifier in the paper. Experiments show that the proposed algorithm can effectively adapt to the target appearance changes, improve the detection accuracy, while maintaining the robustness of a long time tracking.(2) Proposed a pedestrian tracking algorithm based on the deep convolutional neural network(CNN) model. Deep Convolutional neural network can extract features with strong characterization capabilities of the target for images, which can solve the insufficiency and poor robustness of the designed features. The particle filter tracking algorithm is suitable for dealing with nonlinear and non-Gaussian distribution systems. Based on the above theory, we proposed to build the appearance model of the target by a convolutional neural network, and integrated it into the particle filter to track the pedestrian, and update the appearance model by both positive and negative samples generated in the tracking process. By comparing with the state-of-the-art tracking algorithm, we found that the proposed algorithm has good performance in both tracking precision and robustness.