关键词: YOLOv9算法, 多类异常行为检测, 特征提取, 无人机航拍数据集, 深度学习

Abstract: Aiming at the problems of low accuracy of small target recognition and insufficient feature extraction ability in the detection algorithm of UAV aerial images, a multi-category anomaly detection algorithm based on the improved YOLOv9 is designed. The algorithm first adds an improved dimension-sensitive selective integration module to the model head, conducts effective channel segmentation and fusion strategies, and adds a multi-dimensional collaborative attention mechanism to the main body part, while introducing the maximum feature pooling to enhance the feature extraction ability for the self-built dataset. Then, the universal inverted residual module is fused with the original network’s feature extraction module to form the UIB-RepELAN feature extraction module, effectively improving the robustness of the model’s detection. Finally, data augmentation methods are used to expand the anomaly category samples with uneven distribution of easy and difficult samples, and the loss function is reconstructed using Focaler-IoU to improve the model’s generalization ability. The results show that the detection accuracy on the VisDrone2019 dataset is improved from 0.046 to 0.048 compared with the baseline model, and the detection accuracy on the self-built dataset is improved from 0.909 to 0.960, with an average detection time of 28 ms, meeting the requirements of high efficiency and high accuracy detection.

Key words: YOLOv9 algorithm, multi-class abnormal behavior detection, feature extraction, UAV aerial photography datasets, deep learning