A Robust Vehicle Detection Model Based on Attention and Multi-scale Feature Fusion

Vehicle detection plays a critical role in enhancing the environmental perception ability of intelligent vehicles and ensuring safety of vehicles in driving environments. Although many research has been conducted with innovative methods on vehicle detection, improving the robustness of the model in challenging scenes(such as vehicles with a large variance of scales, small-scale vehicles and occluded vehicles) is still an open research question. First, we propose the pyramid multi-scale information module (PMSIM) to collect multi-scale vehicle feature information from dense context, which helps our model adapt to violent changes in vehicle scales. Then, a new detection head for tiny vehicles is added to ameliorate the performance of the model for detecting small-scale vehicles. Finally, we introduce the Transformer encoder to extract global context information to enhance the inference ability of the model. With the help of this ability, the effect of the model to detect occluded vehicles get promoted. The qualitative and quantitative experimental results on KITTI dataset verify that our model improves the robustness and detection accuracy, and outperforming the state-of-the-art methods.