An Improved YOLO Model for Vehicle Recognition System in Aerial Imagery

The modern development in unmanned aerial vehicles (UAV) providing aerial imagery attracts researchers to improve the object detection algorithms to be used in various applications. Lightweight object detection models are required for low computational resource devices. This study developed a lightweight object detection model by improving the architecture of YOLOv4 Tiny to detect vehicles from the VEDAI dataset. In the developed model, one additional scale feature map is added to the architecture. Besides that, the sizes of output images for the second and third prediction boxes are upscaled with the aim of detecting the small pixels of vehicles in the aerial imagery with better accuracy. The experimental results showed an improvement in the detection accuracy and precision when compared with several state-of-the-art methods to detect small objects such as vehicles in aerial imagery.