Small target detection in aerial images of road  scenes based on the improved YOLOv7

Abstract

Small target detection in aerial road images is challenging, and the existing detection models are not suitable for small targets with dense, complicated background and occlusion in aerial images. Therefore, a new small target detection model based on YOLOv7 model is used for small target detection in UAV road scenes. We have made four improvements： (i) a new head for detection is added to the original YOLOv7 model to detect tiny-scale targets. (ii) we use the SimAM attention to focus on areas with small-scale targets. (iii) In the downsampling stage, SPD-Conv model is added to replace the original CNN model to improve the capacity to identify small targets of low resolution in aerial road images. (iv) CARAFE model is introduced to enhance feature fusion and reduce information loss. To demonstrate the feasibility of the model, the first set of contrast experiments were first performed on the Drone dataset, the mAP0.5 value is increased by 8.8% compared with that of the original YOLOv7. Then, experiments are also carried out on the VisDrone2019 dataset. In comparison with the original YOLOv7, the mAP0.5 value of the improved YOLOv7 is increased by 7.2%. This model is equally effective for dense and obscured small targets, our detection performance of the improved YOLOv7 is superior to that of recent methods.