DeepTracking from Aerial Platforms

Object tracking is a pivotal part of video processing. Defense based UAVs generally fly at an altitude of 20,000-30,000ft. Conventionally, it's quite easy to track objects from a stationary camera's feed. Videos from UAVs are hard to process beca
¹
use of abrupt motion of the UAV, jitters, noise and cluttering produced in the video. Tracking and surveillance from such altitudes is quite challenging. This paper illustrates the combination of mod- ified YOLOv4 algorithm with Darknet framework to perform object detection and the DeepSORT tracker to perform object tracking in such challenging environments. The detection model achieves an 11% improvement in Mean Average Precision(mAP) when compared with its predecessors on custom images. When paired with DeepSORT and inferred on a high-end GPU it renders output at 28 FPS, making it suitable for real-time tracking. It is also immune to occlusion and camera movement. The proposed model addresses problems in the existing models that require manual locking of target and high computational complexity. In addition to this, it aims to illustrate automatic tracking of moving objects in real-time.