Determining the Colliding Vehicle in Traffic Accidents Using Hybrid Machine Learning Models

In a world rife with vehicular accidents and traffic incidents, it is known that drivers are more likely than not to shift the blame in an accident rather than admit it. Other than that, there is a noticeable lack of models in the academic sector that allow neural networks to differentiate colliding vehicles from one another and are instead fixated on tracking and detecting traffic accidents as a whole. As such, the researchers propose a way of detecting colliding vehicles and classifying both vehicles as either the ‘colliding’ vehicle or the ‘collided’ vehicle. The processes in this machine learning pipeline are split into three main parts: crash detection—to which the model would use a crash detection algorithm; footage tracking—of which the model would utilise DeepSORT; and lastly a colliding vehicle classification algorithm that uses Gated Recurrent Units (GRUs), all of which will be combined to form a novel machine learning pipeline. The model exhibits very mixed performances when detecting both Vehicle 1 and Vehicle 2 in our testing phase. When detecting Vehicle 1, the model provides a very poor recall and F1-score, meanwhile the detection of Vehicle 2 exhibits a decent amount of precision, recall, and F1-score. Overall, the model provides an accuracy of 42% with a macro average precision of 0.45, a macro average recall of 0.29, and a macro F1-score of about 0.30.