Traffic accidents are one of the significant causes of injury, death, hospitalization and disability. The Road Traffic Injuries Report 2021 by the World Health Organization (WHO) reflects that globally every year, 1.20 million lives are lost as a consequence of road traffic accidents. In addition, between 20 and 40 million more commuters suffer nonlethal wounds, with many sustaining disabilities due to their injury. The rapid development of artificial intelligence and computer vision techniques are generating new opportunities for intelligent traffic and safety systems. In many countries, dashboard cameras (dashcam) are widely used in human operated and autonomous vehicles. A smart and efficient and intelligent system that can anticipate and detect accidents from the dashcam mounted video camera will improve preparedness, prevention and accident management. This paper presents a computer vision-based accident anticipation and detection method. The proposed approach employs a spatial feature based Recurrent Neural Network (RNN) along with Long Short-Term Memory (LSTM) cells to anticipate and detect accidents through dashcam video of vehicles. This method can accomplish accident anticipation about 1.7 seconds prior to its occurring with about 80% recall and 71% precision.

中文翻译：

---

使用深度学习进行基于视觉的事故预测和检测

交通事故是造成伤害、死亡、住院和残疾的重要原因之一。世界卫生组织 (WHO) 发布的《2021 年道路交通伤害报告》显示，全球每年有 120 万人因道路交通事故丧生。此外，还有 20 至 4000 万通勤者遭受非致命伤，其中许多人因受伤而持续残疾。人工智能和计算机视觉技术的快速发展正在为智能交通和安全系统带来新的机遇。在许多国家，行车记录仪广泛应用于人工操作和自动驾驶车辆。智能高效的智能系统可以通过安装在行车记录仪上的摄像机预测和检测事故，从而改善准备、预防和事故管理。本文提出了一种基于计算机视觉的事故预测和检测方法。所提出的方法采用基于空间特征的循环神经网络 (RNN) 以及长短期记忆 (LSTM) 单元，通过车辆行车记录仪视频来预测和检测事故。该方法可以在事故发生前约 1.7 秒完成事故预测，召回率约为 80%，准确率约为 71%。