Deep Learning for Moving Object Detection and Tracking from a Single Camera in UAVs (IROS 2016, Electronic Imaging 2018, Electronic Imaging Best Paper Award, Dataset)
Fig. 1. Results of moving object detection/tracking for a video taken from a UAV. Green and Red bounding boxes represent ground-truth and deep learning detection results, respectively. Our deep learning algorithm is capable of detecting and tracking other small UAVs in the far distance.
Unmanned Aerial Vehicles (UAVs) gain popularity in a wide range of applications from civilian to military purposes. Then, collision avoidance system plays a critical role in operating UAVs in a crowded airspace. Because of cost and weight limitations, UAVs are likely to operate with a single sensor designed to collect imagery. This requires moving object detection/tracking algorithm from a video, which can be run on-board efficiently. Given the sequence of video frames, we estimate background motions via perspective transformation model and then identify salient points in the background subtracted image to detect moving objects. We find spatio-temporal characteristics of each moving object through optical-flow matching and then apply deep learning classifier to detect moving targets which have very different motion/appearance compared with background. We also perform Kalman filter tracking to enforce the temporal consistency on our detection. The algorithm is tested on real videos from UAVs and results in Fig. 1 demonstrate that our algorithm is effective to detect and track small fixed wing UAVs with limited computing resources (Odroid-XU4).