Online Multi-object Tracking Exploiting Pose Estimation and Global-Local Appearance Features

Multi-object tracking is a challenge in intelligent video analytics (IVA) due to possible crowd occlusions and truncations. Learning discriminant appearance features can alleviate these problems. An online multi-object tracking method with global-local appearance features is thus proposed in this paper. It consists of a pedestrian detection with pose estimation, a global-local convolutional neural network (GLCNN), and a spatio-temporal association model. The pedestrian detection with pose estimation explicitly leverages pose cues to reduce incorrect detections. GLCNN extracts discriminative appearance representations to identify the tracking objects, which implicitly alleviates the occlusions and truncations by integrating local appearance features. The spatio-temporal association model incorporates orientation, position, area, and appearance features of the detections to generate complete trajectories. Extensive experimental results demonstrate that our proposed method significantly outperforms many state-of-the-art online tacking approaches on popular MOT challenge benchmark.