Abstract: To precisely identify traffic conflicts involving motor vehicles at urban intersections, a novel automated video detection method based on improved YOLOv8 object detection and DeepSORT trajectory tracking algorithm was proposed. By adding a small-target detection layer, incorporating attention mechanism and optimizing the loss function, the detection performance for small-scale and occluded vehicles was enhanced. An extended Kalman filter (EKF) was employed to handle non-linear motion trajectories, and cubic spline interpolation was utilized to fill in completely missing trajectories, thereby improving trajectory accuracy. Conflicts were quantified based on the time-to-collision (TTC) metric. Research results demonstrate that compared to the YOLOv8 and YOLOv5 algorithms, the improved algorithm has achieved a training accuracy improvement of 6.66% and 8.94%, respectively, and a recall rate improvement of 6.61% and 13.30%, respectively. In terms of tracking performance, compared to YOLOv8+DeepSORT and YOLOv5+DeepSORT, the improved algorithm has achieved an improvement of 4.58% and 7.10% in tracking accuracy, and an increase of 3.82% and 9.49% in tracking success rate, respectively. Based on the conflict detection results of the ROC curve, the improved algorithm achieves an AUC value of 0.854, outperforming other methods.

Key words: transportation engineering; urban intersections; multi-target detection and tracking algorithm; traffic conflict; video recognition technology

摘要： 为精确识别城市交叉口机动车的冲突情况，在改进YOLOv8目标检测和DeepSORT轨迹追踪算法的基础上，提出了一种新的交通冲突视频自动检测方法。通过添加小目标检测层、加入注意力机制及优化损失函数，提升对小尺度及模糊车辆目标的检测性能；采用扩展卡尔曼滤波器(EKF)处理非线性运动轨迹，并利用三次样条插值填补全缺失轨迹，提高轨迹精度；基于碰撞时间(TTC)对指标冲突进行量化。研究结果表明：相较于YOLOv8和YOLOv5算法，文中改进算法的训练精度提升了6.66%、8.94%，召回率提升了6.61%、13.30%；在跟踪性能上，相较于YOLOv8+DeepSORT和YOLOv5+DeepSORT，文中改进算法的跟踪精度提升了4.58%、7.10%，跟踪成功度提升3.82%、9.49%；基于ROC曲线的冲突检测结果，文中改进算法的AUC值达到0.854，优于其它方法。

关键词: 交通工程；城市交叉口；多目标检测追踪算法；交通冲突；视频识别技术