Split Liability Assessment in Car Accident Using Deep Learning

Abstract

Car accidents happen occasionally, and when fortunate, they result in damage to the vehicles rather than personal injuries. However, relief can quickly turn into worry as the involved drivers must determine who is responsible for the damage and to what extent. Split liability assessment in car accidents involves revisiting the collision moment to determine the cause and assign responsibility for property loss. Responsibility is often divided between the drivers, mediated by insurance agents. When both parties cannot agree, legal disputes may lead to court settlements. This situation can result in significant financial and time losses as it requires the involvement of accident experts. In this paper, artificial intelligence technology is employed to mitigate the waste of such societal costs. Split liability assessment in car accidents requires legal interpretation, so we introduced the concept of accident types that can quantitatively determine this. Data collected through web crawling was categorized by accident types and labeled with fault according to the accident type. The dataset for split liability assessment requires handling both temporal and spatial information, as it consists of dashcam footage. To address this, we proposed a network based on 3D CNNs. Additionally, accident videos differ from traditional action recognition tasks in that they present various information arranged over time, rather than repetitive frames. We also proposed a method for aggregating this information. Lastly, to ensure robust performance of the developed artificial intelligence network, experiments were conducted to fine-tune the backbone. We analyzed the results by adjusting the number of layers in the backbone and confirmed the performance based on the pre-trained dataset. Additionally, using the developed network, a GUI program was configured, allowing real-time discussion of the elements on which the network focuses.