DriveWeave v4.0.0 | Forward LiDAR + four cameras

DriveWeave v4.0.0：前向激光雷达与四相机

重新录制高速换道、匝道汇入、匝道驶出、无保护左转、环岛五个 CARLA 原生场景。完整视频按 20 fps 原速播放，前后相机、左右相机和前向点云持续显示，只保留识别标注，不加说明字幕。

采用 CARLA 内置 Model 3 车身代理和量产车风格的参考传感器安装位置，并非特斯拉原厂传感器配置，也没有复刻某款量产车标定。使用官方 Town03/04/05/06 地图定制交通交互；这些场景没有新建道路地图。

四路 RGB 分别运行 YOLOX-S。前视相机与原生 120° 前向 LiDAR 做几何后融合；后方和两侧保持图像检测，不显示不存在的激光点或距离。点云采用实际原生返回。当前为离线预训练检测与几何关联，识别结果未回传控制车辆，没有训练融合网络或跨相机重识别。视频帧率不代表实时推理速度。

发布摘要记录完整来源核对与人工抽样范围。匝道和环岛共四个原生边缘返回超过本项目固定数值校验容差，原始点、严格失败和限定复核均保留，没有裁点或抬高门槛。环岛的短路段采样与入口区域检查范围另见摘要。检测仍有漏框与短时编号切换，视频不构成模型准确率或实车驾驶性能验证。

Bilibili 尚未发布。

DriveWeave v3.1.0 | Camera + LiDAR fusion

Camera 与 LiDAR 在五个完整驾驶场景中持续显示，并加入实际 YOLOX-S RGB 检测、标定点云投影、目标关联与表面测距。相机框和点云中的对应目标使用相同颜色；无足够点云支持时不显示距离。

成片对所有帧和类别统一保留 confidence ≥0.75 的结果。原始0.3门槛预测及其中误检、遮挡误关联的诊断均保留；提高门槛也会减少正确远车的显示。门槛参考了这组录像的事后检查，不能据此报告独立测试准确率。

综合视频为 1920×1080、20 fps、195.05 秒，保留原生车流、Tesla Model 3、跟车和后视画面。没有新增解释字幕、插帧或播放加速。

这是离线预训练图像检测与几何后融合基线，没有训练联合融合模型或接入感知控制闭环。目标真值只用于事后评估，车辆仍由场景规则控制。历史 v3.0 图像框的真值来源保持原说明。

全部 3901 帧具有源帧映射和完整解码检查；额外核对五场景首、中、末 15 帧的 75 个画面区域。人工审阅为抽样，不是全片人工观看。源代码、运行方法与检查范围见 docs/camera-lidar-fusion.md 和 docs/media/accepted-fusion-v3-summary.json。

DriveWeave v3.0.0 — Traffic interactions and native sensors

DriveWeave v3.0.0

本轮重录五个 CARLA 交通场景，增加前后车、邻道车辆、对向车流与环内交通，并加入真实后视相机。车辆使用 CARLA 0.9.16 内置 Tesla Model 3，通过原生车辆物理与公开规则控制运行。

发布包含五条完整跟车视频、五条相机与处理四视图，以及一条综合视频。全部按原速 20 fps 播放，无新增画面文案；综合视频在场景和机位之间剪辑。前视、后视、图像处理及点云处理的独立视频也由公开编码脚本生成。

• 高速：八辆交通角色，慢车跟随、后车接近、约 6.9 秒安全间隙等待及实际换道。
• 匝道：保留完整合流、车道变换和驶出过程，按实际车辆速度与车身净距检查动态间隙。
• 左转：六辆对向车真实通过，主车让行后完成左转和驶出。
• 环岛：四辆环行车、两辆跟随车及另一入口车辆，主车让行、跟车绕行并驶出。
• 传感器：真实 RGB、实例真值区域、LiDAR 地面分离、降采样、几何聚类与标定投影。

源码、可移植场景计划和运行方法见仓库首页及 交通录制说明。原始采集、真实进程退出码、完整视频解码、源图一致性及人工抽样审阅分别记录于 公开验收摘要。

环岛宽区域入口检查的一项原严格 FAIL 保留；独立的真实车道与车身扫掠冲突区域复核通过。左转等待全部对向车辆通过，不声称在车流中穿插；左转与环岛驶出尾段没有持续近距离后车。采样间距检查属于项目验收，不构成道路驾驶安全证明。

图像框采用仿真实例真值，点云采用几何处理；未运行训练后的驾驶或检测模型。本轮新场景为 CARLA 原生录制。综合片最后引用 v2 官方 CARLA/SUMO 协同源录像，其 SUMO GUI 为录制后同步回放。未运行 LimSim。

Native CARLA + SUMO v2 preview: highway and ramps

Native CARLA + SUMO preview

Four complete, untrimmed 20 Hz recordings: highway overtaking/lane changes, ramp entry, ramp exit, and official bidirectional live highway co-simulation. Native Tesla Model 3, RGB, simulator-truth-assisted image processing and native LiDAR processing. The displayed SUMO GUI panes are post-capture synchronized replay, including the independently verified live source. No added captions or HUD.

All four passed full source/motion/action/sensor/video/SUMO audits and actual opening/middle/closing visual sample review. Learned perception models are not used. Wide-junction unprotected left turn and roundabout revisions are still being physically validated and are excluded from this preview release.

Native CARLA + SUMO v2: six complete driving scenarios

Native CARLA + SUMO: complete six-scene v2

Six full native CARLA 0.9.16 recordings use its built-in Tesla Model 3, real RGB/instance cameras and LiDAR. The ego vehicle is already moving in the first recorded frame. Dynamic native traffic, ramp actions, opposing-vehicle yielding at a wider intersection, and roundabout yielding/exit are checked against the actual recorded frames.

Each scene has four complete MP4 files: three native camera/sensor views and one CARLA + native SUMO GUI four-pane view. All retain every 20 Hz source frame at the original speed, with no added titles or HUD. Six-second README GIFs are previews only; the full videos are attached below.

• highway: 334 original frames, 16.7 seconds.
• ramp-entry: 487 original frames, 24.35 seconds.
• ramp-exit: 476 original frames, 23.8 seconds.
• left-turn: 580 original frames, 29.0 seconds.
• roundabout: 1001 original frames, 50.05 seconds.
• live-highway: 334 original frames, 16.7 seconds.

Evidence scope and limits:

• SUMO GUI panes are post-capture replay, including the independently verified live source.
• Simulator-label boxes are ground truth assisted image processing, not learned detection.
• Vehicle separation checks sample recorded world frames, not continuous-time safety guarantees.
• Manual sample review is disclosed separately from full automated video decoding.
• Left-turn SUMO replay retains one empty lane field for a background vehicle at a connected road endpoint (sample 501, native frame 1200430, 0.058394 m from the endpoint). Its complete footprint is outside the actual native GUI viewport. All ego lane fields and all source poses, times and images remain verified; all-actor lane membership is explicitly false. The original strict FAIL is preserved and no lane or pose was filled or modified.
• Roundabout SUMO replay retains 16 empty lane fields: two visible ego samples at connected normal/internal endpoints within 0.1 m and 14 consecutive samples (41-54) of background road44/-5 outside the actual viewport. All-actor and all-ego lane membership are false, while exact full source poses/times and original PNG hashes pass. The original strict FAIL remains; lanes and poses were not repaired. Independent native GUI repeats match actual viewport and all poses and differ by at most 4 pixels; measured PSNR is disclosed, not claimed identical bytes. This is post-capture GUI representation, not a new closed-loop roundabout run.
• Native Traffic Manager routes are requests, not guaranteed path constraints; acceptance uses actual captured motion, body separation and scenario actions.
• The live-highway source separately passed actual official two-way CARLA/SUMO co-simulation checks. Every displayed SUMO GUI pane is explicitly a post-capture synchronized replay.
• Image processing uses native simulator instance labels and RGB edges. LiDAR processing uses actual sensor points, voxel filtering and fitted-ground removal. No trained detector or learned driving policy is claimed.

Cleanup runs after this upload succeeds and all public resources pass full remote byte/SHA-256 verification. It removes only reviewed obsolete generated binaries; source code, maps, metadata and audit history remain.