How High can you Detect? Improved accuracy and efficiency at varying altitudes for Aerial Vehicle Detection

Object detection in aerial images is a challenging task mainly because of two factors, the objects of interest being really small, e.g. people or vehicles, making them indistinguishable from the background; and the features of objects being quite different at various altitudes. Especially, when utilizing Unmanned Aerial Vehicles (UAVs) to capture footage, the need for increased altitude to capture a larger field of view is quite high. In this paper, we investigate how to find the best solution for detecting vehicles in various altitudes, while utilizing a single CNN model. The conditions for choosing the best solution are the following; higher accuracy for most of the altitudes and real-time processing ( >20 Frames per second (FPS) ) on an Nvidia Jetson Xavier NX embedded device. We collected footage of moving vehicles from altitudes of 50-500 meters with a 50-meter interval, including a roundabout and rooftop objects as noise for high altitude challenges. Then, a YoloV7 model was trained on each dataset of each altitude along with a dataset including all the images from all the altitudes, and several training and evaluation experiments were conducted. Overall, the best method for achieving optimal trade-off between accuracy and inference speed is to training on a mixed dataset of multiple altitudes and tune the inference speed to the smaller image size.