Improved YOLOv7 Algorithm for Detecting Occluded Targets at Aircraft Turnaround Milestone

Abstract

The aircraft and special vehicles in the aircraft turnaround milestone have the characteristics of large-scale difference, small spacing distance, etc., and many external interference factors result in many occlusion situations in the process of services. Aiming at the problem that existing airport detection algorithms are ineffective in detecting occluded targets in milestones, an improved YOLOv7 occluded target detection algorithm for aircraft turnaround milestones is proposed. The algorithm adds an efficient channel attention mechanism to the backbone network to suppress the interference information in the environment and improve the attention to the important channels and replaces the convolution in the feature fusion network with a double-pooled coordinate convolution, which is used to improve the model’s ability to extract the spatial information of the occluded target and the remaining features. The algorithm achieves an average detection precision of 89.7% and a recall rate of 88.3% on the aircraft turnaround milestone dataset covering numerous occlusion scenarios, which are 6.1% and 4.4% higher than the YOLOv7 algorithm, respectively. The experiments show that the algorithm can better cope with the task of occlusion target detection in aircraft turnaround milestones.