Online multiple pedestrians tracking using deep temporal appearance matching association

In online multi-target tracking, modeling of appearance and geometric similarities between pedestrians visual scenes is great importance. The higher dimension inherent information in the model compared to problematic many ways. However, due recent success deep-learning-based methods, handling high-dimensional becomes feasible. Among deep neural networks, Siamese network with triplet loss has been widely adopted as an effective feature extractor. Since can extract features each input independently, one update maintain target-specific features. it not suitable for settings that require comparison other inputs. To address this issue, we propose a novel track based on joint-inference network. proposed method enables two inputs be used adaptive contributes disambiguation target-observation matching consolidation identity consistency. Diverse experimental results support effectiveness our method. Our work was recognized 3rd-best tracker BMTT MOTChallenge 2019, held at CVPR2019. (https://motchallenge.net/results/CVPR_2019_Tracking_Challenge/) code available athttps://github.com/yyc9268/Deep-TAMA.