A high-precision vehicle detection model based on focused feature extraction and multi-scale feature fusion for remote sensing images

Abstract

Efficient and accurate vehicle detection is crucial for intelligent transportation and autonomous driving. However, challenges for vehicle detection in complex backgrounds include the low detection accuracy due to insufficient feature extraction and fusion, and high hardware costs as a result of a large number of model parameters. To address these issues, this paper proposes a high-precision vehicle detection model (HVDM). Firstly, through integrating coordinate attention (CA) mechanism into Faster Implementation of Cross Stage Partial Bottleneck with 2 Convolutions(C2f) module, a focused feature extraction module (FFEM) is proposed, which replaces C2f module in backbone to improve the capacity of model for feature extraction. Secondly, triple weighted feature fusion pyramid network (TWFFPN) is constructed by introducing underlying small object feature layer and adding data connections between feature maps of different levels, which improves the ability of multi-scale feature fusion in the network. Finally, normalized Wasserstein distance (NWD) loss is used to substitute for complete intersection over union (CIoU) loss, so the network gets the same sensitivity to objects regardless of their sizes, increasing vehicle detection accuracy. To examine the effectiveness of HVDM, numerous experiments are carried out on datasets for object detection in aerial images (DOTA) and datasets for object detection in optical remote sensing images (DIOR). Experimental results show that our method has a higher accuracy than other methods in vehicle detection and the number of parameters in HVDM is almost the same as that in baseline YOLOv8, which indicates our model's good performance in balancing between detection accuracy and the number of parameters.