Print Email Facebook Twitter Improving Online Multi-Person Tracking Occlusion Title Improving Online Multi-Person Tracking Occlusion: Scale Loss for Deep ReID Feature Learning Author Yang, Hongyu (TU Delft Electrical Engineering, Mathematics and Computer Science) Contributor van Gemert, Jan (mentor) Degree granting institution Delft University of Technology Programme Computer Science Date 2018-11-29 Abstract Occlusion and crossing in Multi-Person Tracking always influence the tracking results. In this paper, we show how deep Re-Identification (ReID), which aims at matching pedestrians across non-overlapping video cameras, can be used to improve the occlusion problem on tracking. The learned ReID feature is more robust than other features used in traditional trackers because the training set is collected from different cameras which includes different parts of the same person.This also helps to solve the occlusion problem in tracking. We train a neural network with the designed scale loss which normalizes both weight vectors and output features to remove the effect of their scale variations on a large Person ReID dataset offline to learn the deep ReID model and build a framework combining detector and tracker to meet real-world application requirements. During the online tracking stage, the data association is solved by calculating the cosine distance cost matrix according to the learned ReID feature vectors. Experiments show that using ReID features can effectively reduce the occlusion index data on MOTChallenge, and the scale loss performs well. Overall our method achieves competitive performance on MOTChallenge, and the framework guarantees the running speed in real-time. Subject Multi-Person TrackingOcclusionScale lossPerson Re-identification To reference this document use: http://resolver.tudelft.nl/uuid:047f6900-3021-459c-9b4b-b5eda4a5c4a1 Part of collection Student theses Document type master thesis Rights © 2018 Hongyu Yang Files PDF Thesis_HongyuYang.pdf 60.94 MB Close viewer /islandora/object/uuid:047f6900-3021-459c-9b4b-b5eda4a5c4a1/datastream/OBJ/view