针对交通碰撞场景下电动两轮车轨迹预测速度慢导致车辆无法快速预测未来碰撞风险的问题,提出一种电动两轮车运动轨迹预测轻量化模型(Tiny Trajectory Prediction,TP-tiny)的建模方法.首先,构建了基于无锚框方法的检测模型(Vehicle and Electric Two-wheelers Detection,VEDet),解决了由于电动两轮车对骑车人遮挡导致的误检漏检问题.同时构建了轻量化跟踪网络(Mobile Simple Online and Realtime Tracking,MobileSort),并采用分组卷积代替传统卷积重新排序卷积层的通道.另外使用Arcface(Additive Angular Margin Loss)为跟踪模型的损失函数,提高了模型在交通碰撞场景下对目标物重识别的能力.其次,使用变分自编码模块学习数据隐变量空间的概率分布潜在属性,并构建新的元素实现对历史轨迹隐数据的降噪处理.采用V-GAN(Generative Adversarial Network with Variational Auto-encoder)预测模块实现电动两轮车未来轨迹预测及图像可视化.最后,基于 TRAF(Dense and Heterogeneous Urban Traffic Dataset)和VRU-TRAVi(Vulnerable Road Users Traffic Accident Database with Video)数据集分别验证了TP-tiny模型的预测性能和运行速度.研究结果表明:VEDet模型大小为2.1 MB,分别为Yolov4-tiny、Yolov5s及CornerNet基准模型的1/6、1/7和1/4,检测速度约为上述基准模型的7.8倍、5.5倍和 1.3 倍;MobileSort(8.6 MB)模型大小(检测速度)分别为 DeepSort、MOTDT 的 1/10(2.25倍)和1/2(2.13倍);TP-tiny模型大小为14.3 MB,运行速度约为Social-GAN、GCN的5.46倍和3.7倍.在未来3 s的预测时域中,TP-tiny模型在保证电动两轮车轨迹预测精度的前提下,进一步提高了预测速度,证明了模型的有效性.
Lightweight Modeling of Electric Two-wheeler Trajectory Predictions in Traffic Collisions
The slow trajectory prediction speeds of electric two-wheelers in traffic collision scenarios leads to the inability of vehicles to quickly predict the risk of a future collision.In response to this issue,this study proposed a modeling method of a tiny trajectory prediction(TP-tiny)for electric two-wheeler motion trajectory predictions.First,Vehicle and Electric Two-wheeler Detection(VEDet),which is based on the anchor-free method,was constructed to address the issue of missed and false detections caused by electric two-wheelers obstructing the rider.A mobile,simple,online,real-time tracking tool called MobileSort was also developed,with group convolution replacing traditional convolution to reorder the channels of the convolutional layer.Moreover,Arcface loss(i.e.,additive angular margin loss)was used as the loss function for the tracking model,thereby enhancing the ability of the model to re-identify targets in traffic collision scenarios.Second,a variational auto encoder module was employed to learn the probabilistic distribution of latent attributes in the hidden variable space of the input data,and new elements were constructed for the denoising of historical trajectory hidden data.The Generative Adversarial Network with Variational Auto-encoder(V-GAN)prediction module was employed for predicting the future trajectories of electric two-wheelers and for image visualization.Finally,the prediction performance and operation speed of the TP-tiny model were validated using the Dense and Heterogeneous Urban Traffic Dataset(TRAF)and Vulnerable Road Users Traffic Accident database with Video(VRU-TRAVi)datasets.The results showed that the VEDet model,with a size of 2.1 MB,was approximately 1/6,1/7,and 1/4 of the sizes of the YOLOv4-tiny,YOLOv5s,and CornerNet benchmark models,respectively,and its detection speed was about 7.8,5.5,and 1.3 times faster than those of the benchmark models,respectively.The MobileSort model(8.6 MB)was 1/10 the size of and 2.25 times faster than DeepSort,and it was the size of and 2.13 times faster than MOTDT.The TP-tiny model,with a size of 14.3 MB,operated approximately 5.46 and 3.7 times faster than the Social-GAN and GCN models,respectively.In the future 3 s prediction time frame,the TP-tiny model further increased the prediction speed while ensuring the accuracy of electric two-wheeler trajectory predictions,thereby proving the effectiveness of the model.
NETL
NSTL
万方数据
