Machine learning-based vehicle detection and tracking based on headlight extraction and GMM clustering under low illumination conditions

• Published in: Expert systems with applications 2025-04, Vol.267, p.126240, Article 126240
• Main Authors: Lashkov, Igor, Yuan, Runze, Zhang, Guohui
• Format: Article
• Language: English
• Subjects: Headlight extraction; Nighttime conditions; Traffic flow monitoring; Vehicle detection; Vehicle tracking
• ISSN: 0957-4174
• DOI: 10.1016/j.eswa.2024.126240

Advanced traffic flow management and control systems aimed at continuously monitoring vehicles are quite popular due to video camera affordability and their wide applicability in intelligent transportation domain. To date, vehicle object detection and tracking at nighttime built upon the use of RGB color sensor-based, in particular, monocular cameras, is a challenging task for researchers and engineers. Light reflections, coming from the vehicle lights, road signs, road surface markers, and headlight’s dazzle may significantly distort the camera image and, eventually, drop the performance of the vehicle surveillance systems. To overcome these issues and build this approach, we employ convolutional neural networks to extract the headlights of each vehicle so as to detect their location at the same time. Once the position of every vehicle is known, we perform a newly made adaptive method built on top of Gaussian Mixture Models clustering based on prior knowledge and assumptions to pair up the extracted single headlights belonging to the same vehicle and track the vehicle objects sequentially frame by frame in a video. The trajectories of the vehicles could be further analyzed to generate the traffic volume and speed parameters for traffic flow control. The experiments conducted in different environmental conditions show stable and robust results for the developed approach. Our proposed method achieved the best overall performance with an average of 97.4 % IDF1, 96.8 % IDR, 94.7 % Recall, 104 MT, 398 FN, and 90.6 % MOTA indicators on a public headlight detection and tracking dataset.