A Robust Model Based on Context Attention and Multi-Scale Feature Fusion for Vehicle Detection

Vehicle detection is a significant part of autonomous vehicles. Although many vehicle detection approaches have achieved impressive performance, achieving robust vehicle detection is still challenging due to the problem caused by occlusion, truncation, small-size vehicles and vehicles with a large variance of scales. In this paper, we propose a vehicle detection model based on the context attention module (CAM) and the multi-scale feature fusion module (MSFFM). Firstly, we construct the backbone of the vehicle detection network with ELAN and MAP. Secondly, we propose the context attention module (CAM) to establish long-range dependencies between pixels, which help the vehicle detection model capture useful context information to promote the robustness of model in challenging scenarios. Finally, we propose the multi-scale feature fusion module (MSFFM) to enhance the model's capability to adapt vehicles with a large variance of scales by collecting multi-scale features. Sufficient qualitative and quantitative experimental results on the KITTI dataset verify that our model achieves the impressive detection accuracy and speed, and outperforming the state-of-the-art methods.