High-altitude Multi-object Detection and Tracking based on Drone Videos

Abstract

Drone videos have more extensive shooting ranges, more angles, and no geographical limitations. Thus the object detection algorithm based on drone videos is increasingly playing a role in various fields, such as military surveillance, space remote sensing, smart city, disaster monitoring scenes, etc. Compared to low-altitude object detection and tracking (LA-ODT), high-altitude object detection and tracking (HA-ODT) are receiving increasing attention, especially in modern cities with massive high buildings, because of their higher flying h eight, w ider v iewing a ngle, a nd t he a bility t o t rack multiple f ast-moving o bjects s imultaneously. However, high-altitude aerial videos (HA-AVs) are constrained by small objects that can be measured, fewer feature points, occlusions, and light changes. Therefore, HA-AVs suffer from blurry images with fewer feature points of objects and missed detection due to occlusion, degrading the ODT accuracy. Since the accessible HA datasets are very limited, not to mention featured datasets considering angles, weather, etc., this paper directly uses drones to collect HA pictures and videos of different angles, different illuminations, and different heights for self-labeling training. Regarding this, we adopt super-resolution reconstruction to increase the data diversity and add artificial o cclusions t o e nhance t he c ollected d ata t o improve t he a ccuracy o f HA-ODT.