Traffic Conflicts Analysis in Penang Based on Improved Object Detection With Transformer Model

Abstract

Current road safety detection and analysis tends to focus on hit-and-run accidents that have already occurred, while ignoring near-misses that may pose a potential safety risk. A monitoring method for near-misses based on the improved YOLOv7 of transformers is proposed in this work. First, the backbone network of YOLOv7 is improved using a G3HN structure ( $g^{n}$ Conv) with recursive gate convolution. Second, the global attention mechanism (GAM) is added to the probe to improve recognition accuracy. Finally, the object detection results are inverted by inverse perspective mapping (IPM) to obtain the centroid of the object, and then the probability of near miss is calculated and analyzed using the DN-based, PICUD-based, and PSD-based methods. This experiment was based on the POL37 Closed-Circuit Television (CCTV) dataset from Penang, Malaysia. The experimental results show that the improved algorithm proposed in this paper can effectively identify small targets in the object detection phase with a detection accuracy of 94.8%, smaller models, and faster training speed, and can fulfill the surveillance task of near-miss events in CCTV surveillance scenes.

Keywords

• G3HN
• near-miss events
• object detection
• transformer mechanism