Attention-based Deep Feature Class Proximity for Detection of Traffic Anomaly in Imbalanced Dataset

期刊论文
发表日期：

2025年9月22日

查看全文

Shrusti Porwal,
Preety Singh,
Anukriti Bansal

摘要

Anomaly detection in traffic surveillance videos is vital for enabling timely responses to incidents that disrupt traffic flow and compromise safety. In this study, we propose a novel framework that combines attention-based deep feature extraction with class distribution proximity for classification of anomaly detection in traffic scenes. To address the severe class imbalance in traffic anomaly datasets, we incorporate an $\alpha $ $α$ -balanced focal loss during training, improving the model’s sensitivity to rare anomalous events. We evaluate our method on the CADP dataset and achieve an F1-score of 0.753 and an AUC of $84.3\%$ $84.3 %$ , marking a significant improvement in the context of highly imbalanced data. We also conduct a comprehensive comparison with other state-of-the-art deep learning models. Our results demonstrate that integrating attention-based features with proximity-based classification and a tailored loss function provides an effective and scalable solution for anomaly detection in real-world traffic surveillance systems.