get step

README.md

@@ -60,8 +60,10 @@ cd /home/zhangw/scan-calibrator
 > 1
 已读取: pan=13420 (134.20°), tilt=-340 (-3.40°), zoom=810 (8.10×)
 
-④ 请保持zoom不变，把摄像头移到扫描区域中央、距离较远且纹理丰富的位置。
-请避开纯天空、纯墙面、水面、玻璃和重复图案；确认后程序会自动移动、拍照并计算step。
+④ 这一步用于测量当前zoom下的水平/垂直视场。
+请保持zoom不变，优先把摄像头移到扫描区域内部、距离较远且纹理分布在整幅画面的位置。
+请避开纯天空、纯墙面、水面、玻璃、近距离物体和重复图案；
+回复1后不要再操作Web云台，程序会自动移动、拍照并返回起点。
 到位后回复 1；回复 3 随时退出。
 > 1
 
@@ -81,8 +83,8 @@ cd /home/zhangw/scan-calibrator
 ⑤ 本次得到的扫图参数：
 {
   "pan_start": 13420,
-  "pan_end": 19672,
-  "pan_step": 1042,
+  "pan_end": 19870,
+  "pan_step": 1075,
   "tilt_start": -340,
   "tilt_end": 1511,
   "tilt_step": 617,
@@ -90,7 +92,7 @@ cd /home/zhangw/scan-calibrator
 }
 安装方式：desktop
 范围检查：视觉左→右pan=decreasing，指定弧段=57.00°
-整步扩边：pan_step=1042，向终点外扩=552（5.52°）
+整步扩边：pan_step=1075，向终点外扩=750（7.50°）
 整步扩边：tilt_step=617，向终点外扩=521（5.21°）
 预计网格：7列 × 4行 = 28张
 
@@ -164,20 +166,25 @@ smart-cam的HTTP接口读取PTZ并记录，等价于：
 
 ### ④ 自动计算pan_step和tilt_step
 
-现场人员把镜头移到扫描区域中央、距离较远且纹理丰富的位置。回复`1`后，向导会：
+现场人员保持zoom不变，把镜头移到扫描区域内部、距离较远且纹理丰富的位置。
+回复`1`后，向导会：
 
 1. 抓取基准图；
 2. 沿pan方向尝试若干候选角度，每次抓图后回到起点；
 3. 用ORB特征、RANSAC和单轴画面位移求水平视场；
 4. 沿tilt方向重复测量；
-5. 按`step = 视场 × (1 - 0.04)`计算目标重叠4%的step；
-6. 要求至少50个几何内点且内点率至少45%，否则拒绝结果；
-7. 删除本次自动标定产生的临时截图。
+5. 每个方向用不同移动量反算视场：两组结果只有在特征质量高且视场差不超过8%时才接受，
+   否则要求至少三组一致结果；
+6. 水平方向按1%目标重叠计算step，垂直方向按4%目标重叠计算step；
+7. 要求至少50个几何内点且内点率至少45%，否则拒绝结果；
+8. 删除本次自动标定产生的临时截图。
 
 这一步不能省略。两个角点只能确定扫描范围，无法确定当前摄像头型号和zoom下的一幅画面
 覆盖多少角度。
 
 不要把画面对准纯天空、纯墙、水面、玻璃、近距离物体或重复图案，否则无法可靠匹配。
+这一步不能改变zoom；自动对焦不影响zoom。标定失败时可以只更换第四步的位置重试，
+不需要重新记录左上角和右下角。
 
 ### ⑤ 显示最终参数
 
@@ -261,5 +268,7 @@ docker exec smart-cam curl -fsS \
 - 每一个摄像头点位都要重新记录角点；
 - 摄像头型号或zoom变化时必须重新自动计算step；
 - 标定期间不要同时从Web界面控制PTZ；
-- 向导目标是4%重叠，为5%上限保留1%误差；
+- 水平目标重叠为1%，垂直目标重叠为4%，都不超过5%上限；
+- 自动标定不会使用单个样本；两组样本必须同时满足高特征质量和视场差不超过8%，
+  普通质量的结果必须由至少三组样本互相验证；
 - 机械回差、镜头畸变和近景视差可能导致不同位置略有差异，最终必须完整试扫并拼图复核；

interactive_calibrator.py

@@ -20,7 +20,7 @@ import scan_calibrator as core
 
 ROOT = Path(__file__).resolve().parent
 CONFIG_PATH = "/smart-cam/config/config.json"
-TARGET_OVERLAP = 0.04
+TARGET_OVERLAPS = {"pan": 0.01, "tilt": 0.04}
 RED = "\033[31m"
 GREEN = "\033[32m"
 YELLOW = "\033[33m"
@@ -67,8 +67,8 @@ def remove_session_snapshots(paths: set[str]) -> None:
 
 def candidate_deltas(axis: str) -> list[int]:
     if axis == "pan":
-        return [250, 450, 700, 1000, 1400, 1900, 2600]
-    return [120, 220, 350, 500, 700, 950, 1300]
+        return [150, 250, 350, 450, 600, 800, 1100, 1500, 2100]
+    return [80, 120, 170, 220, 280, 350, 450, 600, 850, 1150]
 
 
 def choose_direction(axis: str, origin: dict[str, int], points: dict) -> int:
@@ -85,7 +85,7 @@ def estimate_axis(axis: str, origin: dict[str, int], points: dict,
     direction = choose_direction(axis, origin, points)
     baseline = core.snapshot()
     snapshots.add(baseline)
-    reliable: list[tuple[float, int, dict]] = []
+    reliable: list[tuple[int, float, float, dict]] = []
     say(f"\n正在标定 {axis}_step，摄像头会自动移动并回到起点……")
 
     for delta in candidate_deltas(axis):
@@ -102,28 +102,79 @@ def estimate_axis(axis: str, origin: dict[str, int], points: dict,
         say(f"  step={delta}: 重叠约 {overlap*100:.1f}%，"
             f"有效特征={inliers}，内点率={ratio*100:.1f}%")
         if inliers >= 50 and ratio >= 0.45 and 0.08 <= overlap <= 0.80:
-            # Prefer a mid-overlap sample: enough common features without tiny motion.
-            reliable.append((abs(overlap - 0.40), delta, result))
-        if 0.18 <= overlap <= 0.55 and reliable:
-            break
+            reliable.append((delta, overlap, delta / (1.0 - overlap), result))
 
-    if not reliable:
+    if len(reliable) < 2:
         core.move(origin)
         core.fail(
-            f"{axis}方向没有得到可靠视场测量。请把摄像头移到远处、纹理丰富、"
-            "非重复图案的位置后重新运行向导"
+            f"{axis}方向只得到{len(reliable)}个可靠样本，至少需要2个不同移动量互相验证。"
+            "请把摄像头移到远处、纹理丰富、非重复图案的位置后重新运行向导"
         )
 
-    _, measured_step, detail = min(reliable, key=lambda item: item[0])
-    measured_overlap = float(detail["overlap"])
-    estimated_fov = measured_step / (1.0 - measured_overlap)
-    step = max(1, round(estimated_fov * (1.0 - TARGET_OVERLAP)))
+    # Fixed OSD text and repeated scenery can yield a convincing false
+    # homography. A real optical FOV must agree across different PTZ deltas.
+    clusters = [
+        [
+            sample for sample in reliable
+            if abs(sample[2] - seed[2]) / max(seed[2], 1.0) <= 0.12
+        ]
+        for seed in reliable
+    ]
+    consistent = max(
+        clusters, key=lambda cluster: (len(cluster), -max(x[2] for x in cluster))
+    )
+    if len(consistent) < 2:
+        core.move(origin)
+        values = ", ".join(f"{sample[2]:.0f}" for sample in reliable)
+        core.fail(
+            f"{axis}方向不同移动量反算的视场不一致（{values}），疑似重复图案误匹配；"
+            "本次结果不写入配置，请更换标定画面"
+        )
+
+    fovs = sorted(sample[2] for sample in consistent)
+    estimated_fov = (
+        fovs[len(fovs) // 2] if len(fovs) % 2
+        else (fovs[len(fovs) // 2 - 1] + fovs[len(fovs) // 2]) / 2
+    )
+    spread = (max(fovs) - min(fovs)) / estimated_fov
+    if len(consistent) == 2:
+        strong_pair = all(
+            int(sample[3].get("inliers") or 0) >= 100
+            and float(sample[3].get("inlier_ratio") or 0) >= 0.65
+            for sample in consistent
+        )
+        allowed_spread = 0.08
+        if not strong_pair:
+            core.move(origin)
+            core.fail(
+                f"{axis}方向目前只有2个一致样本，但特征质量不足以使用双样本规则；"
+                "需要第3个可靠样本，请更换标定画面"
+            )
+    else:
+        allowed_spread = 0.15
+    if spread > allowed_spread:
+        core.move(origin)
+        core.fail(
+            f"{axis}方向视场样本离散度为{spread*100:.1f}%，"
+            f"超过{allowed_spread*100:.0f}%；"
+            "本次结果不写入配置，请更换标定画面"
+        )
+
+    target_overlap = TARGET_OVERLAPS[axis]
+    step = max(1, round(estimated_fov * (1.0 - target_overlap)))
+    measured_step, measured_overlap, _, detail = max(
+        consistent, key=lambda sample: int(sample[3].get("inliers") or 0)
+    )
     core.move(origin)
+    say(
+        f"  {axis}视场一致性通过：{len(consistent)}个样本，"
+        f"估算视场={estimated_fov:.0f}，离散度={spread*100:.1f}%"
+    )
     return {
         "value": step,
         "direction": direction,
-        "overlap": TARGET_OVERLAP,
-        "target_overlap": TARGET_OVERLAP,
+        "overlap": target_overlap,
+        "target_overlap": target_overlap,
         "overlap_method": "axis_shift_v1_retargeted",
         "zoom": origin["zoom"],
         "measured_at": origin,
@@ -134,15 +185,43 @@ def estimate_axis(axis: str, origin: dict[str, int], points: dict,
             "value": measured_step,
             "overlap": measured_overlap,
             "estimated_fov": round(estimated_fov, 4),
+            "samples": [
+                {
+                    "value": sample[0],
+                    "overlap": round(sample[1], 6),
+                    "estimated_fov": round(sample[2], 4),
+                    "inliers": sample[3].get("inliers"),
+                    "inlier_ratio": sample[3].get("inlier_ratio"),
+                }
+                for sample in consistent
+            ],
+            "spread": round(spread, 6),
         },
     }
 
 
-def calculate_steps(points: dict, snapshots: set[str]) -> None:
+def calculate_steps(points: dict, snapshots: set[str], mount: str) -> None:
     origin = camera_ptz()
     zooms = [int(points[name]["zoom"]) for name in ("top-left", "bottom-right")]
     if max(zooms + [origin["zoom"]]) - min(zooms + [origin["zoom"]]) > 10:
         core.fail("三个位置的zoom不一致；请固定焦段后重新开始")
+    top_left = points["top-left"]
+    bottom_right = points["bottom-right"]
+    pan_start, _, pan_span, _ = core.scan_pan_range(
+        int(top_left["pan"]), int(bottom_right["pan"]), mount
+    )
+    pan_inside = (origin["pan"] - pan_start) % 36000 <= pan_span
+    tilt_low = min(int(top_left["tilt"]), int(bottom_right["tilt"]))
+    tilt_high = max(int(top_left["tilt"]), int(bottom_right["tilt"]))
+    tilt_inside = tilt_low <= origin["tilt"] <= tilt_high
+    if not pan_inside or not tilt_inside:
+        say(
+            "警告：当前标定位置不在刚才选择的扫描区域内。"
+            f"当前pan={origin['pan']}、tilt={origin['tilt']}；"
+            f"扫描区域tilt={tilt_low}～{tilt_high}。"
+            "同zoom下仍可尝试，但优先选择扫描区域内部的远处纹理。",
+            color=YELLOW,
+        )
     state = core.load_state()
     state["steps"]["pan"] = estimate_axis("pan", origin, points, snapshots)
     state["steps"]["tilt"] = estimate_axis("tilt", origin, points, snapshots)
@@ -310,15 +389,26 @@ def main() -> None:
         state["points"]["bottom-right"] = bottom_right
         core.save_state(state)
 
-        confirm_position(
-            "④ 请保持zoom不变，把摄像头移到扫描区域中央、"
-            "距离较远且纹理丰富的位置。\n"
-            "请避开纯天空、纯墙面、水面、玻璃和重复图案；"
-            "确认后程序会自动移动、拍照并计算step。"
-        )
-        say("\n正在计算，请不要操作摄像头或Web控制页面……")
         points = core.load_state()["points"]
-        calculate_steps(points, snapshots)
+        while True:
+            confirm_position(
+                "④ 这一步用于测量当前zoom下的水平/垂直视场。\n"
+                "请保持zoom不变，优先把摄像头移到扫描区域内部、"
+                "距离较远且纹理分布在整幅画面的位置。\n"
+                "请避开纯天空、纯墙面、水面、玻璃、近距离物体和重复图案；"
+                "回复1后不要再操作Web云台，程序会自动移动、拍照并返回起点。"
+            )
+            say("\n正在计算，请不要操作摄像头或Web控制页面……")
+            try:
+                calculate_steps(points, snapshots, mount)
+                break
+            except SystemExit as exc:
+                say(f"\n本次视场标定没有通过：{exc}", color=RED)
+                choose(
+                    "请换一个纹理更丰富的位置重新尝试第四步，两个角点无需重录。\n"
+                    "重新尝试回复 1；结束对话回复 3。",
+                    {"1"},
+                )
         scan, generated, detail = generate(mount)
         say("\n⑤ 本次得到的扫图参数：", color=GREEN)
         say(json.dumps(scan, ensure_ascii=False, indent=2))