Application of Deep Association for Real Time Pedestrian Tracking
Chuan‐Yu Chang, Yuzhang Lin, You-Da Su · 2020
Multiple object tracking plays an important role in computer vision and video analysis. There are many problems with object tracking, such as appearance changes, distance from the camera, occlusion, moving too fast, and so on. In this paper, we combine the pre-trained pedestrian association model with a pedestrian's appearance and moving model to achieve better tracking performance. We trained a neural network base on a large dataset of pedestrian classification, together with the moving model of an object's position, velocity, and acceleration, to help us predict the trajectory more accurately. To demonstrate the performance of the proposed method, the Multiple Object Tracking (MOT) benchmark was used. Experimental results showed the proposed method achieves reasonable tracking results.