YOLO-ME: an enhanced lightweight YOLOv7 Tiny model for efficient object detection in aerial imagery

Vehicle detection in aerial images presents several challenges, including small object sizes, complex backgrounds, varied orientations, and scale variability, all influenced by changing illumination, weather conditions, and high image resolution. Current one-stage detection models have achieved excellent accuracy; however, their complex structures are unsuitable for real-time aerial object detection on embedded devices. To address this issue, this paper proposes the YOLO-ME model, developed based on the YOLOv7 Tiny architecture. The proposed model incorporates three major modifications, including adopting a novel SEELAN module in the backbone, the MELAN module in the neck section, and the adoption of Swish activation functions in most convolutional layers. Experimental results evaluated on the VisDrone dataset show that the YOLO-ME model surpasses other lightweight YOLO-based models, attaining an mAP of 24.82%. Additionally, it features the smallest model size and BFLOPs value at 18.8 MB and 4.68 BFLOPs, respectively. The proposed model demonstrates a detection speed comparable to the YOLOv7 Tiny model. The results indicate that the YOLO-ME model has great potential for real-time aerial object detection on embedded systems. Our code is available at: https://github.com/hanifjunos/YOLO-ME.