Car crash detection using ensemble deep learning

With the recent advancements in Autonomous Vehicles (AVs), two important factors that play a vital role to avoid accidents and collisions are obstacles and track detection. AVs must implement an accident detection model to detect accident vehicles and avoid running into rollover vehicles. At present many trajectories-based and sensor-based multiple-vehicle accident prediction models exist. In Taiwan, the AV Tesla sedan's failure to detect overturned vehicles shows that an efficient deep learning model is still required to detect a single-car crash by taking appropriate actions like slowing down, tracking changes, and informing the concerned authorities. This paper proposes a novel car crash detection system for various car crashes using three deep learning models, namely VGG16(feature extractor using transfer learning), RPN (region proposal network), and CNN8L (region-based detector). The CNN8L is a novel lightweight sequential convolutional neural network for region-based classification and detection. The model is trained using a customized dataset, evaluated using different metrics and compared with various state-of-the-art models. The experimental results show that the VGG16 combined with the CNN8L model performed much better when compared to other models. The proposed system accurately recognizes car accidents with an Accident Detection Rate (ADR) of 86.25% and False Alarm Rate (FAR) of 33.00%.