Spatial-Temporal Analysis for Traffic Incident Detection Using Deep Learning

The domain of Automatic Incident Detection(AID) on roads is predominantly focused on accident detection, lacking the ability to understand specific types of other incidents which may involve accidents. Moreover, existing methods often exhibit limited generalizability, making them susceptible to high error rates when faced with unknown scenarios. To handle these research gaps, this research proposes a deep learning approach that incorporates both the spatial and temporal aspect of videos for the purpose of multiclass video classification in traffic incident detection. By leveraging the modified LRCN network, we enable classification of videos into eight distinct road incident categories. We utilize the Highway incident detection dataset (HWID12) which has more than 2700 videos in different classes. Through rigorous training and comprehensive validation, our approach achieves a precision of 70.30% in the 8-class classification. When considering the dataset as a 2-class classification problem, we achieved an accuracy of 98.02%, which matches the benchmark accuracy on this dataset. These results demonstrate the efficacy of our CNN based framework for automated traffic incident detection, overcoming the limitations of existing methods in terms of specificity and generalizability.