Realtime multi-aircraft tracking in aerial scene with deep orientation network

Abstract

Tracking the aircrafts from an aerial view is very challenging due to large appearance, perspective angle, and orientation variations. The deep-patch orientation network (DON) method was proposed for the multi-ground target tracking system, which is general and can learn the target’s orientation based on the structure information in the training samples. Such approach leverages the performance of tracking-by-detection framework into two aspects: one is to improve the detectability of the targets by using the patch-based model for the target localization in the detection component and the other is to enhance motion characteristics of the individual tracks by incorporating the orientation information as an association metric in the tracking component. Based on the DON structure, you only look once (YOLO) and faster region convolutional neural network (FrRCNN) detection frameworks with simple online and realtime tracking (SORT) tracker are utilized as a case study. The Comparative experiments demonstrate that the overall detection accuracy is improved at the same processing speed of both detection frameworks. Furthermore, the number of Identity switches (IDsw) has reduced about 67% without affecting the computational complexity of the tracking component. Consequently, the presented method is efficient for realtime ground target-tracking scenarios.