A Traffic Flow Monitoring Model Integrating Dual Attention Mechanisms and Densely Connected Architecture

Abstract

The task of monitoring high-density heterogeneous traffic flow presents significant challenges, including insufficient monitoring accuracy, limited adaptability to complex environments, and low computational efficiency. Although current traffic flow monitoring models have improved performance through optimized architectures and lightweight designs, they still face issues such as vulnerability to interference, instability in complex tracking conditions, and inadequate cross-domain generalization. This paper proposes a novel approach that enhances object detection and tracking by integrating attention mechanisms and feature reuse strategies. Specifically, a convolutional block attention module (CBAM) is incorporated into the baseline object detection framework, strengthening feature selectivity and adaptive weighting to enhance robustness under challenging visual conditions. To further refine tracking, the backbone of the DeepSORT algorithm is replaced with a densely connected convolutional network (DenseNet), enabling multilevel feature aggregation and richer target representations, thereby mitigating identity switching and trajectory fragmentation in dense urban traffic. Although the proposed enhancements are demonstrated within a specific object detection and tracking framework, the fundamental contributions—namely attention-enhanced detection and dense feature aggregation—are widely transferable and can be applied across various detection and tracking architectures. Experimental results demonstrate that the proposed method significantly reduces missed detections and improves detection confidence, maintaining reliable target representations under complex traffic conditions with dense, heterogeneous vehicles and dynamic interactions. Ultimately, the proposed framework achieves a practical balance between detection accuracy and real-time performance for traffic flow monitoring.