Risk Prediction of Vehicle Collision Involved with Motorbikes: An Application of A Combined Neural Network of CNN and LSTM

抄録

Advanced Driver Assistance Systems (ADAS) have been implemented to help drivers avoid accidents. ADAS monitors the vehicle's surroundings using data from active sensors. Different data, especially collisions involving motorbikes, are collected to study the potential benefits of ADAS, and deep learning techniques are applied to examine the data. Although the collected data cannot reflect the full features of ADAS, this study can still provide some insights into the integration of ADAS and deep learning techniques. This data is provided by the Tainan City Traffic Accident Investigation Committee, including the video recorded by dashcam or closed-circuit television (CCTV) to simulate the sensor of ADAS and train the risk prediction models to avoid vehicle collision. Deep learning has been widely used as a method of classification and detection in traffic studies. A Convolutional Neural Network (CNN) can capture spatiotemporal dependence through distributed and hierarchical feature extraction. A long short-term memory (LSTM) network can capture the temporal features of videos. This study collects two types of data: static data from accident reports and image data from collected video clips. Five models based on CNN or LSTM are constructed to predict vehicle collisions. The ResNet-50 network, a pre-trained CNN, captures image features from each video frame. LSTM captures the temporal features of videos. F1-score is used to evaluate the performance of models. The results show that integrating CNN and LSTM using vehicle dynamic feature data and video data provides higher performance. Regarding practical applications, if vehicles are equipped with sensors, Models III and V can support ADAS with pre-warning alarms. Drivers or vehicles can respond to these alarms to take appropriate actions to avoid collisions.