Traffic Accident Detection Via Fusion-Gru Based Future Frame Detection Box Prediction

With the rapid development of autonomous driving and advanced driver assistance systems, efforts to enhance driving safety, especially in detecting traffic accidents when natural driving states are unclear, have increased significantly. Generally, traffic anomalies are defined as events that deviate from common occurrences and expected behaviors, such as vehicles and pedestrians disregarding traffic signals to cross arbitrarily, or vehicles abruptly stopping ahead. In real driving scenarios, timely detection of potential traffic accidents by autonomous vehicles is crucial for prompt response and accident prevention. In this work, we aim to address such challenges regarding traffic accident detection by proposing a future frame detection box prediction model based on Fusion-GRU for predicting future vehicle positions. In the evaluation based on two datasets, our results demonstrate that our approach can provide significant improvements in detection performance over similar studies.