Multiple Small-Scale Object Detection in Aerial Vehicle Images using Standard or Optimized YOLO Detectors

Abstract

Widespread employment of low-altitude aerial vehicles that are equipped with cameras and communication devices has increased the range of mobile surveillance applications. Object tracking, anomaly detection, synopsis view construction, or statistical analysis systems using aerial vehicle images require efficient multiple small-scale object detection as an initial stage. In this study, standard YOLO object detectors were trained with the images of VisDrone-2019 using computational resources of a cloud platform, and these detectors were optimized for various hardware systems considering their embedded use in aerial vehicles. Comparisons between six different YOLO versions indicated that YOLOv8-1280 and YOLOv7-1280 were the most precise but the slowest detectors for small-scale objects in aerial images with their 0.489% and 0.396% mean average precision (mAP) values, respectively. However, the computational speeds of YOLOv8-640 and YOLOv7-640 were almost quadrupled without significant precision loss. According to the experiments that compared the standard and optimized versions of YOLOv3, YOLOv4, and YOLOv5, while the optimized detector versions that were generated using TensorRT could process about 2 times more frames per second (fps), up to 6% reduction in their mAP values was observed. Overall, YOLOv5-640 was the best model in trading off the mAP values for fps with its 5.67 and 65.56 fps values in Jetson Nano and Jetson Orin.