transfuser_carla

TransFuser: 模仿基于变压器的传感器融合用于自动驾驶，[cvpr'21] 用于端到端自动驾驶的多模态融合变压器
为了生成基于交叉点的方案的路线，我们检查通过生成的方案有效的交叉点的所有遍历，并包括这些用于训练。在8个公共城镇地图上，我们通过交叉路口获得了大约2500条这样的路线。这些路线的平均长度为100m。由于其他情况，我们的数据集中有大量的交叉路口，因此我们沿着弯曲的高速公路收集场景1和3的数据。大约有500条这样的路线。平均路线长度为400m，并且该场景在路线的中间产生。最后，由于更改车道具有挑战性，因此我们提取了类似于场景1和场景3的弯曲路线，但包括沿路线而不是CARLA场景的车道更改。大约有500条车道变更路线，平均长度为400m。每条这样的路线都有2个车道变更，一个在起点，一个在中途点。

TransFuser: Imitating Transformer-based Sensor Fusion for Autonomous Driving. [CVPR 2021] Multimodal Fusion Transformer for End-to-End Autonomous Driving

To generate routes for intersection-based scenarios, we examined all valid traversals of intersections based on the generated scenarios and included these routes for training. On 8 public town maps, we collected approximately 2,500 such routes via intersections, with an average length of 100 meters per route.

For other scenarios that contain abundant intersections in our dataset, we collected data for Scenarios 1 and 3 along curved highways. There are roughly 500 such routes, with an average length of 400 meters per route, and such scenarios occur in the middle section of the routes.

Finally, given that lane changes are challenging tasks, we extracted curved routes similar to Scenarios 1 and 3 but incorporated lane changes along the routes rather than CARLA scenarios. There are approximately 500 lane-change routes in total, with an average length of 400 meters per route. Each such route includes two lane changes: one at the starting point and the other at the midpoint.

搜集汇总

背景与挑战

背景概述

该数据集基于TransFuser模型，专为自动驾驶设计，包含三种类型的训练路线：基于交叉口的路线（约2500条，平均长度100米）、弯曲高速公路场景（约500条，平均长度400米）和车道变更场景（约500条，平均长度400米，每条包含2个车道变更）。这些数据旨在模拟真实驾驶中的挑战性情况，如交叉口处理和车道变更，以支持端到端自动驾驶模型的训练。