ASSD-YOLO: a small object detection method based on improved YOLOv7 for airport surface surveillance

Wentao Zhou, Chengtao Cai, Liying Zheng, Chenming Li, Daohui Zeng

Abstract: The airport surface surveillance system is essential in ensuring airport safety and maximizing the efficient utilization of airport resources. Current airport detection algorithms suffer from few and lack relevant airport target data. To solve these issues, this paper establishes two airport datasets named ASS-Dataset, including the surveillance dataset and the panoramic surveillance dataset. Compared to other aircraft datasets, our datasets are collected from authentic airport surface surveillance systems. According to observation, most objects are small in datasets. This paper proposes a small object detection method ASSD-YOLO based on improved YOLOv7. First, the designed f-efficient attention module is added to the backbone network to improve the accuracy of the algorithm. Second, the transformer encoder network is incorporated into the backbone network to increase feature extraction. Finally, the small target detection layer is added to the head network to improve the ability to extract small targets. The model of the mean average precision is 93.5\(\%\) in the surveillance dataset. In the panoramic surveillance dataset, the ASSD-YOLO achieves 10.8\(\%\) and 21.4\(\%\) higher average precision for the airplane and truck than YOLOv7. g. Comparing the method proposed in this paper to the original YOLOv7, the performance improvement for the mAP is 4.6\(\%\) when using the RSOD open dataset. ASS-Dataset is available at https://github.com/rookie257/rookie257.github.io. Code is available at https://github.com/rookie257/small_detection.github.io.