A Novel Variant of Yolov7-Tiny for Object Detection on Aerial Vehicle Images

With the advancement of deep learning and the rising popularity of unmanned aerial vehicle (UAV) surveillance systems, object detection from UAV aerial imagery has gained attention in computer science. Numerous high-accuracy solutions have partially solved the problem of detecting small objects. Most current approaches use deep learning models with large sizes and many parameters, making them challenging to deploy to edge devices. Besides, some others focus on reducing model size but not maintaining accuracy. This study proposes a novel model based on Yolov7-tiny to address the abovementioned problems. The solution will simultaneously meet multiple criteria, including accuracy, number of parameters, and model capacity. Our proposed model achieved the best performance with 50% mAP.50 in the Visdrone2019-DET dataset, 5.5 M parameters, 11.8 MB of model weights, and 57 FPS of inference time on Nvidia RTX 3090. Moreover, the experimental results confirm that our method outperforms the Edge-Yolo.