Vehicle identification from surveillance camera

• Main Author: Yang, Nanyang
• Other Authors: Lap-Pui Chau
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2021
• Subjects: Engineering::Computer science and engineering::Computing methodologies::Image processing and computer vision; Engineering::Electrical and electronic engineering
• Online Access: https://hdl.handle.net/10356/149935
• Institution: Nanyang Technological University

Computer vision has been a popular research topic, especially with the advanced development of deep learning models and improved hardware support. Common tasks like object detection and tracking can be used to improve work efficiency and security. For this final year project, we focused on machine learning task multi-object tracking (MOT), particularly from surveillance cameras. This technique can be utilised to monitor traffic conditions and be part of the smart city transportation solutions for efficient traffic management. However, many MOT methods suffer from high identity switches due to the inferior association method. In this project, we proposed a tracked object bounding box association based on the CenterTrack algorithm with tracked object size change thresholding. We conducted ablative studies on the MOT17 challenge training dataset and evaluated our proposed method on the MOT17 test set and UA-DETRAC vehicle tracking dataset. Our proposed method CenterTrack++ can reduce identity switches significantly by 22.6% and obtain a notable improvement of 2.3% in IDF1 tracking score compared to the original CenterTrack’s under the same tracklet lifetime on the MOT17 test dataset, achieving the best performance among the trackers only using spatial features in the association. Evaluations on UA-DETRAC also yield similar results, with a significant reduction of 14.7% and 41.9% in identity switches compared to the original CenterTrack and the state-of-the-art FairMOT tracker. Furthermore, we developed a real-time traffic monitoring dashboard to demonstrate one possible application of the proposed method on self-collected traffic surveillance videos, using our proposed association method. The research paper on the CenterTrack++ association method has been accepted for publication at the 2021 ICME International Workshop on Big Surveillance Data Analysis and Processing.