基于深度学习的智能交通车流监测与预测研究

Research on Intelligent Traffic Flow Monitoring and Prediction Based on Deep Learning

为了方便交通部门改善交通拥堵问题,使用旭日X3嵌入式开发板作为硬件平台,通过YOLOv8深度学习网络识别道路上通行的车辆及其车辆类型。使用开放神经网络交换(Open Neural Network Exchange, ONNX)格式可视化编辑工具去掉了模型的输出头,将网络中的激活函数由SiLU函数更换为ReLU函数,将模型输出由80个检测类别更改为4个检测类别,在Small版本中,使用非极大值抑制算法(Non-Maximum Suppression, NMS)将最合适的检测框筛选出来,然后用SORT(Simple Online and Realtime Tracking)多目标追踪算法和匹配算法将独立帧检测到的车辆关联起来,实现车辆自动计数。在服务器上配置好YOLOv8的训练环境,训练3个周期,测试模型的mAP指标为0.635,推理速度提升至20 fps左右,目标检测系统的计数精度达到98%,可以准确获取到路口的交通流数据,帮助改善交通拥堵问题。

In order to facilitate the transportation department to improve traffic congestion,this paper proposes to use the Horizon Sunrise X3 embedded development board as a hardware platform to identify vehicles and their models on the road through the YOLOv8 deep learning network.The Open Neural Network Exchange(ONNX)format visual editing tool is used to remove the output header of the model,replace the activation function SiLU in the network with ReLU function,and change the model output from 80 detection categories to 4.In version Small,Non-Maximum Suppression(NMS)algorithm is used to filter out the most suitable detection boxes.Then,the vehicles detected by independent frames are associated with the multi-target tracking algorithm and matching algorithm using SORT(Simple Online and Realtime Tracking)to realize automatic vehicle counting.The training environment of YOLOv8 is configured on the server with a training period of 3 cycles.With the mAP index of the test model being 0.635 and the reasoning speed increasing to about 20 fps,the counting accuracy of the target detection system reaches 98%,which can accurately obtain the traffic flow data at the intersection and help improve the traffic congestion problem.