Real-Time Detection of Anomalous Driving Behavior Using Generative Adversarial Networks and Behavior Prediction Models
Daoyi Xiao, Yulu Hu, Xinyi Sun · 2024
In light of the accelerated advancement of autonomous driving technology, the assurance of the system's security and dependability has emerged as a pivotal concern. One of the principal challenges in this regard is the real-time detection of abnormal driving behaviour, which is essential for the avoidance of accidents and system failures. In this work, we propose a real-time detection framework based on a Generative Adversarial Network (GAN) and a behaviour prediction model with the objective of identifying abnormal behaviours in autonomous vehicles. Generative adversarial networks are employed to ascertain the distributions of typical driving behaviour and to generate realistic driving data. Meanwhile, behaviour prediction models are utilised to anticipate the anticipated behaviour of the vehicle based on historical sensor inputs and environmental context. The specific methods include the use of a generative network to simulate normal driving trajectories and a discriminant network to identify the differences between the real and generated data. Furthermore, the behaviour prediction model employs a more sophisticated deep learning architecture to capture intricate driving behaviour patterns and to compare them with real-time observed behaviours. When a significant deviation is identified, an anomaly alert is triggered. The efficacy of the model in detecting anomalous driving behaviours with high precision and low latency has been validated through a substantial number of experiments conducted on both simulated and real-world driving datasets.