Changes

Update mizar.jenkinsfile

Update polaris.jenkinsfile

Update tc397_ota.jenkinsfile

Update hdmap-cloud.jenkinsfile

Update map-png.jenkinsfile

feat(trt10): 引入 BUILD_TRT10 编译开关, 适配 x86 Blackwell + TRT10 环境

CMakeLists 增加 option(BUILD_TRT10), 默认 OFF:
  - OFF (默认): GPU 架构 = 原 SMS 列表 (30 32 35 37 50 52 53 60 61 62 70 72 75 87),
                行为与改动前完全一致, Orin/Xavier/老 x86 部署不受影响.
  - ON : GPU 架构 = sm_120 (RTX 50xx Blackwell), 同名宏 BUILD_TRT10 自动注入
         C/C++ 和 NVCC, 子模块通过 add_subdirectory 继承.
         源码侧走 TRT10 新 API (delete / setMemoryPoolLimit / enqueueV3 +
         setTensorAddress / setInputShape 等), 见 centerpoint/pointpillars
         submodule 同名 commit.

子模块指针更新:
  - centerpoint  -> 068e2dd (cp_lib)
  - pointpillars -> af67324 (dev)

同时纳入本机 RViz 视图配置: conf/perception_imv.rviz.

使用:
  catkin build                                       # 老环境, 行为不变
  catkin build --cmake-args -DBUILD_TRT10=ON         # 本机 Blackwell+TRT10

Co-authored-by: Cursor <cursoragent@cursor.com>

note_out:车挂角度校正添加箱体面估计新算法.

note_out:车挂角度校正箱体面算法修改.

note_out:车挂角度校正日志和显示修改.

note_out:前箱面由小格选单点变为选一定范围内的点.

note_out:车挂校正无可靠初始角度时搜索挂前沿侧面计算粗糙参考角度.

note_out:车挂校正无可靠初始角度时搜索挂前沿侧面计算粗糙参考角度,适配小角度计算方法.

note_out:车挂校正无可靠初始角度时搜索挂前沿侧面计算粗糙参考角度,适配小角度计算方法.

note_out:恢复误删除的车挂显示相关topic&&拆分dld_trailer_calibrator.cpp中两个类到两个单独的文件.

note_out:修改enable_expand_filter_用法.

note_out:增加各来源角度框的可视化.

note_out:无参考角度初始角度估计优化.

note_out:无参考角度时车挂角度搜索第二套方法.

note_out:修复无车周点时NoiseFilter车周栅格不更新空置状态的bug.

note_out:过滤岸桥轨道上的沟沿误检.

note_out:第二套无参考初始挂角查找优化.

trailer front-edge: fix large-angle misclassify + logging + micro-opt

- classify: replace |dx|vs|dy| small-angle heuristic with two-hypothesis
  (along/normal) picked by |y_tr| closeness to half-width; kills spurious
  +90deg at large trailer angles
- add z-band gate to reject cargo-box side face (out of baffle height band)
- selection by |y_tr|-to-half-width instead of longest extent
- angle-compare log: initial_baffle / initial_front_edge / revised / ekf / detect
- micro-opt: consider() takes precomputed fit (drop duplicate LS on single
  segment path); reserve cell vectors; std::move cell idx

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

trailer front-edge batch1: box-face channel + z-classification + fusion

Treat box/body front face (tall grid cells above baffle height band) as a
second, parallel estimator of the same trailer angle instead of discarding it:

- consider_box(): route runs whose z-top exceeds the baffle band to a
  box-face candidate; long straight wall -> dominant-axis classify
  (normal at small angle, along at large), keep the longest wall
- fusion: prefer baffle (on-target height); fall back to box-face angle only
  when baffle route finds nothing (e.g. loaded trailer, baffle occluded) ->
  recovers cases that previously abstained (a7)
- logging: per-candidate cand-box line; fuse line with th_baffle/th_box/src/out;
  RESULT now tags src=baffle|boxface
- viz/out use fused final_a/final_b/final_side

NOTE: not compile-verified in this environment (needs project Docker build).

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

trailer front-edge batch2: raw-point fit + dual-side pooling + y-geo + confidence

- fit_raw(): LS over raw points of the chosen segment (not the 4 grid
  centroids) to suppress the per-column max-x quantization bias that drove
  the ~-7deg systematic error at small angle (a4/a6)
- dual-side pooling: gather baffle-height cells L/R, sum per-side de-meaned
  covariances -> shared face direction (does NOT connect the two centroids,
  so it is immune to the cross-side motion-compensation x offset); used to
  refine short baffle segments
- y-position geometric fallback: cosθ=(cyL-cyR)/(2w) from known half-width,
  y-only (avoids polluted x); logged (sensitive at large angle only)
- confidence scalar per source; logged in fuse/RESULT
- new logs: batch2 pool (nL/nR/th_pool/th_geo_mag), batch2 refine
  (th_centroid/th_raw/th_pool/use_pool)

Behavior change (intended): baffle final angle now comes from raw-point fit,
or dual-side pool for short segments.
NOTE: not compile-verified in this environment (needs project Docker build).

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

note_out:宝世顺去铁轨矮草散点.

note_out:烟台高柱检出区对行人标记为地图外.

note_out: 矮柱近邻查找增加质心距离约束.

note_out: 矮柱近邻查找质心距离约束0.2->0.25.

note_out:修改高柱栅格标记时域确认方式.

note_out:修改高柱栅格标记时域确认方式.

note_out:高柱历史信息记录标记使用时间戳而非帧计数.

note_out:无初始角度适配小角度.

note_out:删除第一种求初始角度方法.

note_out:初始角度查找.

note_out:修改暂存.

龙须点里不满足细线的点抛弃

note_out:烟台增加立体车库防护功能:高墙面拟合输出110点.

龙须点线粗细约束&&x向和z向的离群值剔除

note_out:大榭像素42区域低矮障碍物点加入102点.

note_out:无初始角度时估计车挂初始角度v3最终版.

note_out:调节箱面角度的置信度.

note_out:双侧挡流板面拟合角度值相互校验条件放松.

note_out:车挂日志打印条件修改.

note_out:恢复仅无初始车挂角时才调用周围搜索.

note_out: 挡流板置信度校验去除score1,2逻辑&&代码优化.

note_out: 挡流板置信度校验去除score1,2逻辑&&代码优化.

note_out:参数修改.

note_out:增加箱体面计算挂角的注释.

note_out:拟合改为pca,增加箱体面角度y向箱宽校验.

note_out:pca去噪点改为单次.

note_out:调试日志,topic清理和隐藏.

note_out:箱面半宽调整.

note_out:坑洼检测地面分割后处理: 突变点检测,地面地下点粗标签标记.

围栏点也经龙须点筛选标记

参数控制是否使用“收紧龙须点”的规则，洋浦不收紧

note_out: 优化地面分割后处理:聚类和分类逻辑优化.

note_out: seg坑沿点辨识优化.

note_out:seg判坑簇修改条件.

note_out:增加梁山后向雷达输出显示topic.

note_out:魏桥og点柳树过滤.

note_out:魏桥og点柳树过滤.

note_out:willow_hang_z默认改成4.5, 防误用.

note_out:增加dl输出角度的保存和可视化输出.

note_out:堆高机正面吊取消速度纠正box heading.

note_out:铁路堆场路沿点增加时域叠加功能.

note_out:地磅/沟沿/路沿检出车身屏蔽区提取为参数.

note_out:new_detection的地图显示增加像素相关着色&&增加map系密集栅格等间距格点云输出&&增加魏桥柳树矮枝像素和过滤方法.

note_out:魏桥柳树区不做未定义点捞回.

nogte_out:去掉调试代码.

note_out:路口和道路内35米内不触发box refiner.

note_out:振石窄且短地磅时域叠加策略.

note_out:删除调试时的错误map_origin赋值.

note_out:前港cp更新135k模型,更新submodule model指针[新增sweeper/sprinkler/car类别].

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

note_out:前港新模型更新,洒水车和清扫车新类别适配.

note_out: 铁路堆场出口像素53添加过滤和稳定逻辑.

note_out:路沿时域叠加端点排查日志(注释版).

note_out:extra NROI加高度范围参数.

note_out:宝世顺铁路堆场路沿ROI过滤参数(curb_filter_roi_utm,借像素53过滤逻辑).

新增curb_filter_roi_utm(UTM多边形)参数,在像素53过滤位置对落在ROI内的路沿点一并过滤不输出.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

note_out:下一时刻过滤框可视化.

note_out:烟台车内点聚类初步实现.

note_out:烟台自车内聚类优化.

note_out:接入载车状态消息,控制是否开启自车内聚类和112点标记输出.

note_out: 107点的z最低阈值恢复0.1米.

note_out:梁山更新专有模型.

note_out:前港集卡车头高分阈值0.4->0.55.

转向角一个超阈值则整组去框

feat(replay_script): 部署脚本相对路径化 + 驱动目录env注入 + 克隆脚本

- run_driver.sh: 配置走相对路径；驱动靠 SETUP_<pkg> 环境变量 source（无硬编码/无搜索），
  hesai 外参用 rospack find 推导；新增 --dry-run/--background
- run_perception.sh: 相对化，自动上溯定位 share/perception，/senior 收敛为 SENIOR_ROOT，新增 --dry-run
- 新增 clone_new_detection.sh（加固版 worktree 克隆）、deploy.env.example、.gitignore、README
- driver_config.yaml: imv_4pxt32_2rsbp4 改指非bp4 imv launch（BP3/BP4自动判型合用）
- CMakeLists: 安装 tools/replay_script 到 share/perception/tools/replay_script

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

docs(replay_script): README 补充感知安装路径设置，更新过时注意项

- 快速开始新增“感知安装路径”步骤（PERCEPTION_SHARE/PERCEPTION_SETUP/SENIOR_ROOT）
- run_perception.sh 段落说明非 install 内运行需显式设 PERCEPTION_SHARE
- 环境变量表标注感知安装/系统根相关项
- 注意项更新：2 个 launch 已推送；改列缺失的 5 个感知档分支

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

note_out:前港洒水车和清扫车类别稳定策略.

note_out:烟台新过滤策略.

note_out:烟台新加过滤策略先触发录包搜集信息.

梁山洗车间谁水花过滤

调整洗车间过滤距离阈值