Improving vehicle detection accuracy in complex traffic scenes through context attention and multi-scale feature fusion module

Abstract

Vehicle detection is a fundamental task for automated driving systems. However, achieving robust performance in complex traffic scenarios remains a formidable challenge. In this paper, we propose a novel vehicle detection model that leverages contextual attention mechanisms and multi-scale feature fusion to effectively tackle the inherent challenges presented by complex scenarios, such as occlusion, truncation, and small-scale vehicle instances. The proposed model introduces a contextual attention module tailored to address vehicle occlusion, augmenting the model’s reasoning ability and overall performance through the integration of global contextual information. Additionally, we introduce a Multi-Scale Feature Fusion Module to mitigate the impact of drastic changes in vehicle scales observed in dynamic traffic scenarios. Through the deployment of a dedicated multi-scale feature fusion module, our model adeptly adapts to significant size variations of vehicles in traffic scene images, thereby enhancing its capability to handle vehicles of varying sizes. Our contributions are validated through comprehensive qualitative and quantitative experiments conducted on both the KITTI dataset and the Cityscapes dataset. The experimental results demonstrate the exceptional robustness and accuracy of our proposed model. These findings provide conclusive evidence of the superior performance and effectiveness of our model in real-world applications.