基于人机协同共融的车辆轨迹预测模型

本团队基于Carla模拟器构造了一个端到端的自动驾驶框架作为轨迹预测算法验证平台。该平台包括物体识别模块、轨迹预测模块和运动控制模块。首先，以相机、Lidar等传感器数据为输入，利用卷积神经网络识别交通灯状态、车辆、行人等。然后，以论文中提出的ON-LSTM模型为轨迹预测模型，通过捕捉驾驶员的动态记忆数据，预测车辆未来的轨迹。最后，以车辆轨迹为输入，控制自动驾驶车辆安全通过测试场景。Carla模拟器（开源地址：https://carla.org/）记录车辆轨迹数据。考虑到驾驶环境的复杂性，汇交的数据集涵盖了城市、学校、郊区、乡村和高速公路等15类场景以及跟随、变道、转弯及超车等任务的驾驶数据。

Our team developed an end-to-end autonomous driving framework based on the Carla simulator as a verification platform for trajectory prediction algorithms. This platform comprises three core modules: object recognition, trajectory prediction, and motion control. First, leveraging sensor data from cameras, Lidar and other perception devices as input, convolutional neural networks (CNNs) are deployed to recognize traffic light states, vehicles, pedestrians and other road participants. Next, the ON-LSTM model proposed in this work is adopted as the trajectory prediction module, which predicts the future trajectories of vehicles by capturing the dynamic memory patterns of drivers. Finally, taking vehicle trajectories as input, the autonomous driving system is controlled to safely navigate through test scenarios. The Carla simulator, open-sourced at https://carla.org/, records vehicle trajectory data during experiments. Given the complexity of driving environments, the compiled dataset covers 15 types of scenarios including urban, campus, suburban, rural, highway and others, as well as driving data for tasks such as car-following, lane changing, turning and overtaking.

搜集汇总

背景与挑战

背景概述

该数据集是基于Carla模拟器构建的车辆轨迹预测模型数据集，包含城市、学校、郊区、乡村和高速公路等15类场景的驾驶数据，涵盖跟随、变道、转弯及超车等任务。数据集由清华大学团队创建，数据量为14.72MB，包含36个文件。