Enhanced YOLOv11m for Real-Time Multi-Scale Traffic Detection under Haze Conditions

Abstract

In intelligent transportation systems, accurate traffic object detection plays a vital role in improving road safety, traffic monitoring, and autonomous driving applications. However, detecting vehicles, pedestrians, and other road objects under hazy weather conditions remains a challenging task due to reduced visibility, low contrast, blurred object boundaries, and background noise. Existing lightweight object detection models such as YOLOv11n provide real-time performance but often suffer from reduced detection accuracy in adverse weather environments, especially when dealing with multi-scale objects and complex traffic scenes.

To address these limitations, this project proposes an enhanced object detection model named Proposed v11m, developed based on the YOLOv11n framework. The proposed system introduces several architectural improvements to enhance feature extraction and multi-scale representation while maintaining lightweight computational complexity. Firstly, an Attention-Gate Convolution (AGConv) module is integrated into the backbone network to improve contextual awareness and suppress irrelevant background information. Secondly, a Multi-Dilation Sharing Convolution (MDSC) module is incorporated to capture features at multiple receptive fields, improving sensitivity toward objects of varying scales. Additionally, a Cross-Channel Feature Fusion Module (CCFM) is designed within the neck network to adaptively recalibrate channel-wise feature importance and strengthen feature fusion.

The proposed model is trained and evaluated using traffic datasets containing haze-affected road scenes. Experimental results demonstrate that Proposed v11m achieves improved detection accuracy compared to the existing YOLOv11n model, with higher mAP values while preserving real-time inference speed. The model achieves efficient performance with only 2.6 million parameters and high frame processing capability, making it suitable for deployment on resource-constrained edge devices and embedded intelligent traffic systems.

Overall, the Proposed v11m model provides an effective and robust solution for real-time traffic object detection in challenging hazy environments, contributing toward safer and more reliable intelligent transportation systems.