Using deep learning and accident report analysis to determine culpability and detect car damage

Conventional car bumper and driver culpability prediction rely on the experience of specialised experts. A long wait may result from obtaining such a prediction at the accident scene. It is possible to avoid this kind of wait by generating an automatic report. In this paper, a culpability estimation approach based on an analysis of accident reports and detection of damage caused in cars is presented. Our system employs a fine-tuning-based approach to enable the prediction of the kind, position, and intensity of vehicle bumps. For culpability prediction, is based on mining data from accident reports. We obtained a precision of 94.7%, a recall of 96.3%, and a mAP of 97.7%. The dataset size for car damage is 7,402 images. As for the prediction of culpability, it gave an accuracy of 91%. The dataset size for accident reports is 135,764 records.