High‐accuracy vehicle lamp detection for real‐time night‐time traffic surveillance

Abstract

Vehicle lamps are an important image feature of the night‐time vehicle detection algorithm. This study proposes a real‐time night‐time vehicle detection algorithm based on light attenuation characteristic analysis, which consists of vehicle lamp detection and pairing. For the detection phase, this study proposes an automatic dual‐threshold method to quickly extract the attenuation regions around bright objects. This method is highly adaptable and can accurately extract attenuation regions to identify vehicle lamps and strong reflections. For the pairing phase, this study uses the moving trajectories of vehicle lamps to complete preliminary pairing. Then, the vehicle lamp pairing is optimised using the relative vertical and horizontal lamp positions in continuous images. Pixel‐ and object‐based methods were used to verify the experimental vehicle lamp detection results. The results of both verification methods indicate that the proposed algorithm without morphological operation is superior to other algorithms. The accuracy of the pairing phase was >94.6% in scenes with multiple lamps, high‐driving speed, high traffic flow, and rain. Finally, the proposed night‐time vehicle detection algorithm can achieve a real‐time execution speed for images 960 × 720 pixels in size.

Research on night-time vehicle target detection based on improved KSC-YOLO V5

“…Hui Li [9] and colleagues achieved an 87.22% detection rate by isolating and matching the features of vehicle front and rear lamps, although this method's reliance on headlight features alone may reduce its accuracy for smaller vehicles. Wen-Kai Tsai et al [10] developed an innovative automatic dual thresholding technique for efficiently extracting the dimming region around bright objects, significantly improving the accuracy of such as R-CNN [11] , Fast R-CNN [12] , alongside Single Stage Detectors like SSD [13] and the YOLO series [14][15][16] . The essence of two-stage detection lies in initially identifying candidate regions, followed by precise location regression and classification within these areas, segmenting the process into distinct stages.Guanxiang Yin et al [17] Zheng et al [18] developed a modification for the Faster R-CNN framework tailored for detecting faint targets in complex traffic scenes.…”

MC-YOLO-Based Lightweight Detection Method for Nighttime Vehicle Images in a Semantic Web-Based Video Surveillance System

“…Compared to other detection technologies (Zhao et al, 2022;Ding et al, 2022;, semantic Web-based video surveillance has several advantages, including convenient installation and maintenance, no need to interrupt traffic, low cost, large amounts of analyzable information, and no impact on road life. Currently, this technology is experiencing rapid development (Tsai & Chen, 2021;Mo et al, 2022;.…”