A-YOLO: small target vehicle detection based on improved YOLOv5

In this paper, we introduce an enhanced YOLOv5-based model tailored to address challenges in detecting small vehicle targets within high-altitude surveillance contexts. Within the backbone network, the SE attention mechanism is incorporated after each convolutional module. To amplify the detection capabilities, a high-resolution detection head is integrated. Within the intermediate (neck) network, we apply context aggregation and feature fusion techniques before the small target detection head, while employing the CBAM attention mechanism for other detection heads. To further optimize performance, the loss function has been revised by substituting the traditional Ciou calculation with Focal Eiou, in tandem with enhancing object localization using the OTA algorithm. Additionally, the Retinex algorithm is deployed for data augmentation. Performance assessments on the VisDrone2019 dataset revealed that our proposed approach outperformed the standard YOLOv5, 2019 challenge sota, and TPH-YOLO (with prior +CBAM) models by margins of 20%, 7%, and 3%, respectively. A series of tests affirmed the robustness of our algorithm, particularly in the realm of small vehicle target detection.