Traffic monitoring and management systems are vital for the maintenance of urban mobility and safety. The accurate and real-time recognition of various traffic elements, such as vehicles, pedestrians, and traffic signs, is crucial for the effective functioning of these systems. However, traditional object detection models often struggle to maintain high accuracy and speed in complex traffic scenarios due to varying lighting conditions, occlusions, and the diverse nature of traffic participants. In this paper, we present TrafficYOLO, an enhanced version of the state-of-the-art object detection model YOLOv4, specifically tailored for traffic detection scenarios. TrafficYOLO incorporates a novel multi-head attention mechanism which allows the model to focus on multiple relevant features simultaneously, thereby improving detection precision and recall in challenging environments. Our multi-head attention module is seamlessly integrated into the YOLOv4 architecture, which already achieves impressive speed and accuracy trade-offs. The multi-head attention module enhances the model's ability to handle occlusions and recognize small or partially visible objects that are common in traffic scenes. Experiments on standard traffic datasets demonstrate that Traffic Yolosignificantly outperforms existing methods in terms of both detection accuracy and real-time processing speed. Traffic Yolois shown to be particularly effective in detecting densely packed objects and in conditions with low visibility, making it a robust solution for real-world traffic monitoring applications.
TrafficYOLO: YOLO with Multi-Head Attention Mechanism for Traffic Detection Scenarios
29.03.2024
1438435 byte
Aufsatz (Konferenz)
Elektronische Ressource
Englisch
Traffic flow prediction method based on multi-head attention mechanism
Europäisches Patentamt | 2023
|