Abstract: Unmanned aerial vehicle (UAV) videos, an important video resources captured by unmanned aerial vehicles, are now being widely used in ground target monitoring. However, there’s usually a large amount of space-time redundancy in UAV videos due to their grand view and unspecified targets, making the traditional methods of video interaction inefficient to get usable details. To solve the problem, a multi-scale spiral summarization for UAV videos was proposed. Firstly, we trained a detection model based on YOLOv3 algorithm to detect the small targets including pedestrians and vehicles from the UAV’s perspective. Then, we proposed a key-frame-based video object detection algorithm, by first extracting the key frames of the videos according to the improved color-feature-based key-frame-extraction algorithm, and then applying the model on the key frames to get the target detection results of the whole video. The key areas from the key frames were extracted as the displaying units of video summarization in a spiral form from the inside out with basic functions including key-frame-based video location and dynamic scaling. At last, some novel extended interaction tools were developed including sketch annotation, object distribution spiral and double spiral player, aiming to meet the users’ potential needs, and help them interact with the UAV videos more efficiently.

Key words:  unmanned aerial vehicle, video summarization, video object detection, small object detection, spiral summarization, video interaction