使用多模型特徵進行駕駛員異常行為檢測

Abstract

近年來，道路交通事故的風險正在迅速上升。異常駕駛仍然是交通事故的主 要原因之一。駕駛員異常行為檢測是一個重要的計算機視覺問題，可以在提高交 通安全和減少交通事故方面發揮至關重要的作用。基於卷積神經網絡(CNN)的 廣泛方法已被應用於駕駛員異常駕駛的檢測。在卷積神經網絡 (CNN) 中，卷積運 算擅長提取局部特徵，但難以捕獲全局表示。為解決這一問題，我們提出了一種 新的駕駛員異常行為的檢測方法，該方法將不同的 CNN 模型以及 ViT 模型相結 合來捕獲局部和全局特徵。此外，特徵融合的過程可以大大增強局部特徵的全局 感知能力和全局表示的局部細節。在 StateFarm 數據集上進行的大量實驗表明，我們提出的方法表現出最佳性能。

The risk of road accidents has been rising rapidly in recent years. Abnormal driving is still one of the main causes of traffic accidents. Driver abnormal behavior detection is an important computer vision problem that can play a crucial role in improving traffic safety and reducing traffic accidents. A wide range of methods based on convolutional neural networks (CNNs) have been applied to the detection of abnormal driving by drivers. In convolutional neural networks (CNNs), convolution operations are good at extracting local features, but struggle to capture global representations. To address this problem , we propose a novel driver abnormal behavior detection method that combines different CNN models and ViT models to capture local and global features. In addition, the process of feature fusion can greatly enhance the global awareness of local features and the local details of the global representation. Extensive experiments on the StateFarm dataset show that our proposed method exhibits the best performance.