AeroDetect: Real-Time Small Object Detection in Aerial Images

Real-time small object detection in aerial images is a challenging task in computer vision because of the complex environmental conditions and fewer pixels occupied by the objects. Along with this getting real-time inference makes this much more difficult. To tackle these challenges we propose an improvement in the classical YOLOv3 network which is capable of removing noise from the images and performing small object detection to while maintaining a high frame frequency. We are using a drone to capture real-time footage from the AIT campus, which is further used for classifying objects like vehicles, pedestrians, etc. Our proposed model “AeroDetect” was able to beat various state-of-the-art models like YOLOv3 and YOLOv5 in various parameters by having higher mAP, and a lesser number of layers and parameters while maintaining a high frame rate which is crucial for our use case. We achieved mAP of 0.402 on the Visdrone dataset as compared to 0.4 and 0.38 by YOLOv3 and YOLOv5, respectively. On the VEDAI dataset, we got mAP of 0.632, whereas YOLOv3 and YOLOv5 achieved 0.578 and 0.604, respectively.