Effective traffic density recognition based on ResNet-SSD with feature fusion and attention mechanism in normal intersection scenes

• Main Authors: Zhang, Qiang, Fu, Yuguang
• Other Authors: School of Civil and Environmental Engineering
• Format: Article
• Language: English
• Published: 2025
• Subjects: Engineering; Object detection; Traffic density recognition
• Online Access: https://hdl.handle.net/10356/182788
• Institution: Nanyang Technological University

In normal intersection scenes, there are many tasks that rely on the recognition of traffic density, such as adaptive traffic signal control and driving risk detection. Traditional methods for traffic density recognition are difficult to use, expensive to deploy, and/or may cause damage to the road surface. In tasks related to traffic density recognition, accurately detecting multiple objects in traffic videos, including those of different classes and small sizes, can be crucial. Notably, the detection of these objects in traffic videos can pose additional challenges, leading to a reduction in the accuracy of traffic density recognition. This study presents an applicable method for traffic density recognition based on deep residual network-single shot multi-box detector (ResNet-SSD) with feature fusion and attention mechanism. In this method, we adopt the deep residual network for feature extraction. Regarding the presented feature fusion structure, it can be employed to integrate feature information and enhance the representation of shallow feature maps. In addition, the squeeze-and-excitation network can be adopted. Finally, we conduct the experiments to verify the performance of our presented method. Regarding the traffic density recognition, our presented method has achieved the accuracy of 0.885 and the latency of 12 ms. Our presented method has excellent performance in handling varying traffic objects, particularly small-sized objects. The significant advantages over traditional methods mitigate issues related to poor portability and potential missing crucial information. And our presented method is verified to be applicable for traffic density recognition in normal intersection scenes.