2026.6.12小更新
This commit is contained in:
@@ -55,4 +55,13 @@ class PID:
|
||||
"""
|
||||
self.last_err = 0.0
|
||||
self.integral = 0.0
|
||||
self.last_time = time.time()
|
||||
self.last_time = time.time()
|
||||
|
||||
def reset_integral(self):
|
||||
"""
|
||||
仅清空积分项(I term),保留上一帧误差以维持微分(D term)连续性
|
||||
|
||||
用于卡尔曼 coast(滑行预测)期间:
|
||||
只靠比例 (P) + 微分 (D) 稳住,积分项清零防止过冲。
|
||||
"""
|
||||
self.integral = 0.0
|
||||
@@ -0,0 +1,621 @@
|
||||
"""
|
||||
target_tracker.py — 轻量级多目标追踪过滤器
|
||||
|
||||
【解决的问题】
|
||||
1. 目标短暂丢帧(2-5帧)导致无人机抽搐 → 卡尔曼滤波 coast/predict 平滑过渡
|
||||
2. 多目标 ID 飘移导致甩头 → 最邻近欧氏距离关联锁定 ID 不变
|
||||
|
||||
【核心设计】
|
||||
- 恒定速度卡尔曼滤波器(Kalman Filter × 8状态:cx, cy, w, h, vx, vy, vw, vh)
|
||||
- 最邻近贪婪关联,阈值 200px 内视为同一目标
|
||||
- 死区预测(Coast/Predict):连续丢失 ≤ max_lost_frames 帧时用滤波器预测值
|
||||
- 超过 max_lost_frames → 释放轨道,对外判丢失
|
||||
|
||||
【典型用法】
|
||||
tracker = TargetTracker(max_lost_frames=8)
|
||||
|
||||
# 每帧将 YOLO 原始检测喂入
|
||||
result = tracker.update(yolo_detections, frame_shape=(480, 640))
|
||||
|
||||
if result["tracked"]:
|
||||
cx, cy = result["center"] # 平滑后的中心点(像素坐标)
|
||||
pid.update(...) # 正常 PID 控制
|
||||
else:
|
||||
pid.reset(); send_zero_velocity() # 判丢 → 刹车
|
||||
|
||||
【依赖】
|
||||
- numpy(与推理模块共用,无新增重量级依赖)
|
||||
- 纯 Python,无 OpenCV / scipy 依赖
|
||||
"""
|
||||
|
||||
import numpy as np
|
||||
import time
|
||||
import logging
|
||||
import copy
|
||||
|
||||
logger = logging.getLogger(__name__)
|
||||
|
||||
# ================================================================
|
||||
# 单目标卡尔曼滤波器
|
||||
# ================================================================
|
||||
|
||||
class _KalmanBoxFilter:
|
||||
"""
|
||||
单目标恒定速度卡尔曼滤波器
|
||||
|
||||
状态向量 (8) : [cx, cy, w, h, vx, vy, vw, vh]
|
||||
位置 + 尺寸 + 各自速度
|
||||
|
||||
观测向量 (4) : [cx, cy, w, h]
|
||||
|
||||
所有坐标使用原始像素值(而非归一化),
|
||||
与 YOLO postprocess 输出的坐标空间一致。
|
||||
"""
|
||||
|
||||
# ---- 噪声权重(经验调优值,适用于 640×480 画面) ----
|
||||
_STD_POS = 1.0 / 50 # 位置过程噪声
|
||||
_STD_VEL = 1.0 / 200 # 速度过程噪声
|
||||
_INI_POS = 2.0 # 初始位置协方差系数
|
||||
_INI_VEL = 100.0 # 初始速度协方差系数
|
||||
|
||||
def __init__(self):
|
||||
dt = 1.0 # 帧间步长(帧率变化时仍用 dt=1,Q 已含尺度因子)
|
||||
|
||||
# ---- 状态转移矩阵 F ----
|
||||
self.F = np.eye(8, dtype=float)
|
||||
for i in range(4):
|
||||
self.F[i, i + 4] = dt # 位置 += 速度 × dt
|
||||
|
||||
# ---- 观测矩阵 H ----
|
||||
self.H = np.zeros((4, 8), dtype=float)
|
||||
self.H[0, 0] = 1.0
|
||||
self.H[1, 1] = 1.0
|
||||
self.H[2, 2] = 1.0
|
||||
self.H[3, 3] = 1.0
|
||||
|
||||
# ---- 协方差初始值 ----
|
||||
self.P = np.eye(8, dtype=float) * self._INI_POS
|
||||
for i in range(4, 8):
|
||||
self.P[i, i] = self._INI_VEL
|
||||
|
||||
# ---- 状态向量 ----
|
||||
self.x = np.zeros(8, dtype=float)
|
||||
self._initialized = False
|
||||
|
||||
# ------------------------------------------------------------------
|
||||
# 公开接口
|
||||
# ------------------------------------------------------------------
|
||||
|
||||
def initiate(self, measurement):
|
||||
"""
|
||||
用首次观测初始化滤波器
|
||||
|
||||
Args:
|
||||
measurement: [cx, cy, w, h] — 像素坐标
|
||||
"""
|
||||
cx, cy, w, h = measurement
|
||||
self.x = np.array([cx, cy, w, h, 0, 0, 0, 0], dtype=float)
|
||||
# 初始协方差:位置置信度中等,速度置信度低
|
||||
self.P = np.eye(8, dtype=float) * self._INI_POS
|
||||
for i in range(4, 8):
|
||||
self.P[i, i] = self._INI_VEL
|
||||
self._initialized = True
|
||||
|
||||
def predict(self):
|
||||
"""
|
||||
预测步骤
|
||||
|
||||
Returns:
|
||||
[cx, cy, w, h] — 预测的像素坐标,未初始化时返回 None
|
||||
"""
|
||||
if not self._initialized:
|
||||
return None
|
||||
|
||||
# ---- 动态过程噪声 Q(与当前目标尺寸相关) ----
|
||||
std_pos = [self._STD_POS * self.x[2], # w 为宽度
|
||||
self._STD_POS * self.x[3], # h 为高度
|
||||
self._STD_POS * self.x[2],
|
||||
self._STD_POS * self.x[3]]
|
||||
std_vel = [self._STD_VEL * self.x[2],
|
||||
self._STD_VEL * self.x[3],
|
||||
self._STD_VEL * self.x[2],
|
||||
self._STD_VEL * self.x[3]]
|
||||
|
||||
q_diag = [s * s for s in std_pos] + [s * s for s in std_vel]
|
||||
Q = np.diag(q_diag)
|
||||
|
||||
# ---- 预测 ----
|
||||
self.x = self.F @ self.x
|
||||
self.P = self.F @ self.P @ self.F.T + Q
|
||||
|
||||
return self.x[:4].copy()
|
||||
|
||||
def update(self, measurement):
|
||||
"""
|
||||
更新步骤
|
||||
|
||||
Args:
|
||||
measurement: [cx, cy, w, h] — 观测到的像素坐标
|
||||
"""
|
||||
if not self._initialized:
|
||||
return
|
||||
|
||||
z = np.array(measurement, dtype=float)
|
||||
|
||||
# ---- 动态观测噪声 R(与目标尺寸相关) ----
|
||||
std_pos = [self._STD_POS * abs(self.x[2]),
|
||||
self._STD_POS * abs(self.x[3]),
|
||||
self._STD_POS * abs(self.x[2]),
|
||||
self._STD_POS * abs(self.x[3])]
|
||||
R = np.diag([s * s * 2.0 for s in std_pos]) # ×2 增强平滑
|
||||
|
||||
# ---- 卡尔曼增益 ----
|
||||
S = self.H @ self.P @ self.H.T + R
|
||||
K = self.P @ self.H.T @ np.linalg.inv(S)
|
||||
|
||||
# ---- 后验更新 ----
|
||||
y = z - self.H @ self.x # 残差(innovation)
|
||||
self.x = self.x + K @ y
|
||||
self.P = (np.eye(8) - K @ self.H) @ self.P
|
||||
|
||||
def get_prediction(self):
|
||||
"""返回预测的观测值 [cx, cy, w, h](未初始化时返回 None)"""
|
||||
if not self._initialized:
|
||||
return None
|
||||
return self.x[:4].copy()
|
||||
|
||||
def get_state(self):
|
||||
"""返回完整状态向量(未初始化时返回 None)"""
|
||||
if not self._initialized:
|
||||
return None
|
||||
return self.x.copy()
|
||||
|
||||
@property
|
||||
def initialized(self):
|
||||
return self._initialized
|
||||
|
||||
|
||||
# ================================================================
|
||||
# 跟踪轨道(Tracklet)
|
||||
# ================================================================
|
||||
|
||||
class _Tracklet:
|
||||
"""
|
||||
一条跟踪轨道,包装 _KalmanBoxFilter + 辅助状态
|
||||
|
||||
生命周期:
|
||||
创建(首次检测到)→ 跟踪中(可能 coast)→ 超过丢失上限 → 回收
|
||||
"""
|
||||
|
||||
__slots__ = (
|
||||
"track_id", "kf", "lost_count", "max_lost",
|
||||
"confidence", "cls_id",
|
||||
"age", "hits", "last_center", "last_box",
|
||||
)
|
||||
|
||||
def __init__(self, track_id, measurement, confidence, cls_id, max_lost=8):
|
||||
"""
|
||||
Args:
|
||||
track_id: 全局唯一轨道ID
|
||||
measurement: [cx, cy, w, h]
|
||||
confidence: 检测置信度
|
||||
cls_id: 类别ID
|
||||
max_lost: 最大连续丢失帧数
|
||||
"""
|
||||
self.track_id = track_id
|
||||
self.kf = _KalmanBoxFilter()
|
||||
self.kf.initiate(measurement)
|
||||
|
||||
self.lost_count = 0
|
||||
self.max_lost = max_lost
|
||||
self.confidence = confidence
|
||||
self.cls_id = cls_id
|
||||
self.age = 1 # 已存在帧数
|
||||
self.hits = 1 # 成功匹配次数
|
||||
|
||||
cx, cy, w, h = measurement
|
||||
self.last_center = (float(cx), float(cy))
|
||||
self.last_box = None # 首次调用 predict/update 后设置
|
||||
|
||||
def predict(self):
|
||||
"""预测 -> 更新 last_center / last_box"""
|
||||
pred = self.kf.predict()
|
||||
if pred is not None:
|
||||
cx, cy, w, h = pred
|
||||
self.last_center = (float(cx), float(cy))
|
||||
self.last_box = self._center_to_box(cx, cy, w, h)
|
||||
self.age += 1
|
||||
return self.last_center, self.last_box
|
||||
|
||||
def update(self, detected_box, confidence, cls_id):
|
||||
"""
|
||||
用检测结果更新轨道
|
||||
|
||||
Args:
|
||||
detected_box: [x1, y1, x2, y2]
|
||||
confidence: 检测置信度
|
||||
cls_id: 类别ID
|
||||
"""
|
||||
cx = (detected_box[0] + detected_box[2]) / 2.0
|
||||
cy = (detected_box[1] + detected_box[3]) / 2.0
|
||||
w = detected_box[2] - detected_box[0]
|
||||
h = detected_box[3] - detected_box[1]
|
||||
|
||||
self.kf.update([cx, cy, w, h])
|
||||
self.last_center = (float(cx), float(cy))
|
||||
self.last_box = detected_box
|
||||
|
||||
self.confidence = confidence
|
||||
self.cls_id = cls_id
|
||||
self.lost_count = 0
|
||||
self.hits += 1
|
||||
self.age += 1
|
||||
|
||||
def mark_miss(self):
|
||||
"""标记当前帧未匹配到检测"""
|
||||
self.lost_count += 1
|
||||
self.age += 1
|
||||
|
||||
@property
|
||||
def is_dead(self):
|
||||
"""超过最大丢失帧数时标记为死亡,可回收"""
|
||||
return self.lost_count > self.max_lost
|
||||
|
||||
@property
|
||||
def is_coasting(self):
|
||||
"""当前帧是否处在预测(无实测)状态"""
|
||||
return self.lost_count > 0
|
||||
|
||||
@staticmethod
|
||||
def _center_to_box(cx, cy, w, h):
|
||||
"""[cx, cy, w, h] → [x1, y1, x2, y2]"""
|
||||
half_w = w / 2.0
|
||||
half_h = h / 2.0
|
||||
return [cx - half_w, cy - half_h, cx + half_w, cy + half_h]
|
||||
|
||||
def __repr__(self):
|
||||
return (f"_Tracklet(id={self.track_id}, "
|
||||
f"lost={self.lost_count}/{self.max_lost}, "
|
||||
f"age={self.age}, hits={self.hits})")
|
||||
|
||||
|
||||
# ================================================================
|
||||
# 多目标追踪器
|
||||
# ================================================================
|
||||
|
||||
class TargetTracker:
|
||||
"""
|
||||
轻量级多目标追踪器
|
||||
|
||||
职责:
|
||||
- 接收每帧原始 YOLO 检测列表
|
||||
- 通过卡尔曼滤波 + 最邻近关联维持稳定目标ID
|
||||
- 提供平滑后的目标框 / 中心点供 PID 控制使用
|
||||
- 自动回收超过丢失上限的旧轨道
|
||||
|
||||
典型配置(建议在实例化时按需调整):
|
||||
| 参数 | 默认值 | 说明 |
|
||||
|------|--------|------|
|
||||
| max_lost_frames | 8 | 目标连续消失帧数上限(~250ms @ 30fps) |
|
||||
| max_association_dist | 200 | 关联距离阈值(像素),超过视为新目标 |
|
||||
| min_hits | 3 | 新轨道确认前需要连续匹配帧数 |
|
||||
|
||||
update() 返回值:
|
||||
{
|
||||
"tracked": bool — 是否存在有效主目标
|
||||
"primary_id": int — 主目标 ID(None 如果无目标)
|
||||
"center": (cx,cy) — 平滑后中心像素坐标
|
||||
"box": [x1,y1,x2,y2] — 平滑后边界框
|
||||
"confidence": float — 当前置信度(coast 期间从最近测量继承)
|
||||
"is_predicted": bool — True = 当前帧 coasting(无实测)
|
||||
"lost_frames": int — 连续未命中帧数
|
||||
"n_active": int — 当前活跃轨道总数
|
||||
"raw": [...] — 关联到的原始检测(coast 时为 None)
|
||||
}
|
||||
或丢失时:
|
||||
{
|
||||
"tracked": False,
|
||||
"primary_id": None,
|
||||
...
|
||||
"lost_frames": > max_lost_frames
|
||||
}
|
||||
"""
|
||||
|
||||
def __init__(self, max_lost_frames=8, max_association_dist=200.0,
|
||||
min_hits=3):
|
||||
"""
|
||||
Args:
|
||||
max_lost_frames: 最大连续丢失帧数(默认 8 ≈ 250-300ms)
|
||||
max_association_dist:关联距离阈值(像素,默认 200px)
|
||||
min_hits: 新轨道确认前最少连续匹配帧数(默认 3)
|
||||
"""
|
||||
self.max_lost_frames = max_lost_frames
|
||||
self.max_association_dist = max_association_dist
|
||||
self.min_hits = min_hits
|
||||
|
||||
self._next_id = 0
|
||||
self._tracks = {} # {track_id: _Tracklet}
|
||||
self._frame_count = 0
|
||||
|
||||
# 缓存上一帧的 primary 结果(锁外创建副本供控制线程读取)
|
||||
self._last_result = {
|
||||
"tracked": False,
|
||||
"primary_id": None,
|
||||
"center": None,
|
||||
"box": None,
|
||||
"confidence": 0.0,
|
||||
"is_predicted": False,
|
||||
"lost_frames": 0,
|
||||
"n_active": 0,
|
||||
"raw": None,
|
||||
}
|
||||
|
||||
# ------------------------------------------------------------------
|
||||
# 主接口
|
||||
# ------------------------------------------------------------------
|
||||
|
||||
def update(self, detections, frame_shape=None):
|
||||
"""
|
||||
每帧调用一次,更新追踪状态
|
||||
|
||||
Args:
|
||||
detections: YOLO postprocess 输出的检测列表
|
||||
[[x1, y1, x2, y2, conf, cls_id], ...]
|
||||
允许为空列表 [](表示无检测)
|
||||
frame_shape: (height, width) 用于边界钳制,可选
|
||||
|
||||
Returns:
|
||||
dict — 主目标跟踪结果(格式见类文档)
|
||||
"""
|
||||
self._frame_count += 1
|
||||
|
||||
# ---- 步骤 1:所有活跃轨道先 predict ----
|
||||
for tid in list(self._tracks.keys()):
|
||||
track = self._tracks[tid]
|
||||
track.predict()
|
||||
|
||||
# ---- 步骤 2:关联 — 提取检测中心 ----
|
||||
if detections:
|
||||
# 提取检测中心 + 尺寸
|
||||
det_centers = [] # [(idx, cx, cy, w, h, conf, cls_id)]
|
||||
for i, det in enumerate(detections):
|
||||
x1, y1, x2, y2, conf, cls_id = det
|
||||
cx = (x1 + x2) / 2.0
|
||||
cy = (y1 + y2) / 2.0
|
||||
w = x2 - x1
|
||||
h = y2 - y1
|
||||
det_centers.append((i, cx, cy, w, h, conf, cls_id))
|
||||
|
||||
# 关联
|
||||
matches, unmatched_tids, unmatched_dids = self._associate(det_centers)
|
||||
else:
|
||||
matches = []
|
||||
unmatched_tids = list(self._tracks.keys())
|
||||
unmatched_dids = []
|
||||
|
||||
# ---- 步骤 3:更新匹配的轨道 ----
|
||||
for tid, did in matches:
|
||||
_, cx, cy, w, h, conf, cls_id = det_centers[did]
|
||||
raw_box = [cx - w/2, cy - h/2, cx + w/2, cy + h/2]
|
||||
self._tracks[tid].update(raw_box, conf, cls_id)
|
||||
|
||||
# ---- 步骤 4:未匹配的轨道 → mark_miss / 回收 ----
|
||||
for tid in unmatched_tids:
|
||||
if tid in self._tracks:
|
||||
self._tracks[tid].mark_miss()
|
||||
if self._tracks[tid].is_dead:
|
||||
logger.debug("轨道 %d 已死亡 (lost=%d), 回收",
|
||||
tid, self._tracks[tid].lost_count)
|
||||
del self._tracks[tid]
|
||||
|
||||
# ---- 步骤 5:未匹配的检测 → 创建新轨道 ----
|
||||
for did in unmatched_dids:
|
||||
_, cx, cy, w, h, conf, cls_id = det_centers[did]
|
||||
measurement = [cx, cy, w, h]
|
||||
new_track = _Tracklet(
|
||||
track_id=self._next_id,
|
||||
measurement=measurement,
|
||||
confidence=conf,
|
||||
cls_id=cls_id,
|
||||
max_lost=self.max_lost_frames,
|
||||
)
|
||||
# 新轨道直接使用测量值作为 last_box
|
||||
new_track.last_box = [cx - w/2, cy - h/2, cx + w/2, cy + h/2]
|
||||
self._tracks[self._next_id] = new_track
|
||||
self._next_id += 1
|
||||
|
||||
# ---- 步骤 6:选取主目标(优先级:置信度 > 命中数 > 贴近画面中心) ----
|
||||
result = self._select_primary(frame_shape)
|
||||
self._last_result = result
|
||||
return result
|
||||
|
||||
def get_last_result(self):
|
||||
"""获取上一帧的跟踪结果(线程安全,外部可直接读取)"""
|
||||
return self._last_result
|
||||
|
||||
def reset(self):
|
||||
"""重置所有轨道(清空追踪状态)"""
|
||||
self._tracks.clear()
|
||||
self._next_id = 0
|
||||
self._frame_count = 0
|
||||
self._last_result = {
|
||||
"tracked": False,
|
||||
"primary_id": None,
|
||||
"center": None,
|
||||
"box": None,
|
||||
"confidence": 0.0,
|
||||
"is_predicted": False,
|
||||
"lost_frames": 0,
|
||||
"n_active": 0,
|
||||
"raw": None,
|
||||
}
|
||||
logger.info("TargetTracker 已重置")
|
||||
|
||||
# ------------------------------------------------------------------
|
||||
# 属性
|
||||
# ------------------------------------------------------------------
|
||||
|
||||
@property
|
||||
def active_tracks(self):
|
||||
"""返回当前所有活跃轨道(dict: {track_id: _Tracklet} 的浅复制)"""
|
||||
return dict(self._tracks)
|
||||
|
||||
@property
|
||||
def active_count(self):
|
||||
"""当前活跃轨道数"""
|
||||
return len(self._tracks)
|
||||
|
||||
# ------------------------------------------------------------------
|
||||
# 内部方法
|
||||
# ------------------------------------------------------------------
|
||||
|
||||
def _associate(self, det_centers):
|
||||
"""
|
||||
最邻近贪婪关联
|
||||
|
||||
Args:
|
||||
det_centers: [(idx, cx, cy, w, h, conf, cls_id), ...]
|
||||
|
||||
Returns:
|
||||
(matches, unmatched_tids, unmatched_dids)
|
||||
"""
|
||||
if not self._tracks or not det_centers:
|
||||
return [], list(self._tracks.keys()), list(range(len(det_centers)))
|
||||
|
||||
# 提取所有轨道的预测位置
|
||||
track_preds = {} # {tid: (cx, cy)}
|
||||
for tid, track in self._tracks.items():
|
||||
cx, cy = track.last_center
|
||||
track_preds[tid] = (cx, cy)
|
||||
|
||||
# 贪婪匹配:每条轨道找最近的未分配检测
|
||||
matches = []
|
||||
assigned_dets = set()
|
||||
assigned_tracks = set()
|
||||
|
||||
# 先将轨道按置信度排序(更可靠的轨道优先匹配)
|
||||
sorted_tracks = sorted(
|
||||
self._tracks.items(),
|
||||
key=lambda kv: kv[1].confidence,
|
||||
reverse=True
|
||||
)
|
||||
|
||||
for tid, track in sorted_tracks:
|
||||
if tid in assigned_tracks:
|
||||
continue
|
||||
pred_cx, pred_cy = track_preds[tid]
|
||||
|
||||
best_dist = self.max_association_dist
|
||||
best_did = None
|
||||
for did, _, d_cx, d_cy, _, _, _ in det_centers:
|
||||
if did in assigned_dets:
|
||||
continue
|
||||
dist = np.hypot(pred_cx - d_cx, pred_cy - d_cy)
|
||||
if dist < best_dist:
|
||||
best_dist = dist
|
||||
best_did = did
|
||||
|
||||
if best_did is not None:
|
||||
matches.append((tid, best_did))
|
||||
assigned_dets.add(best_did)
|
||||
assigned_tracks.add(tid)
|
||||
|
||||
unmatched_tids = [tid for tid in self._tracks if tid not in assigned_tracks]
|
||||
unmatched_dids = [did for did in range(len(det_centers)) if did not in assigned_dets]
|
||||
|
||||
return matches, unmatched_tids, unmatched_dids
|
||||
|
||||
def _select_primary(self, frame_shape=None):
|
||||
"""
|
||||
从活跃轨道中选出主目标
|
||||
|
||||
优先级:
|
||||
1. 含有实测(非 coasting)且确认(hits ≥ min_hits)
|
||||
2. 实测轨道中置信度最高
|
||||
3. 若无实测轨道,取 coast 中最接近画面中心的
|
||||
|
||||
Returns:
|
||||
dict — 与 update() 返回值格式一致
|
||||
"""
|
||||
if not self._tracks:
|
||||
return {
|
||||
"tracked": False,
|
||||
"primary_id": None,
|
||||
"center": None,
|
||||
"box": None,
|
||||
"confidence": 0.0,
|
||||
"is_predicted": False,
|
||||
"lost_frames": 0,
|
||||
"n_active": 0,
|
||||
"raw": None,
|
||||
}
|
||||
|
||||
# 分组:实测轨道 vs coasting 轨道
|
||||
measured = []
|
||||
coasting = []
|
||||
|
||||
for tid, track in self._tracks.items():
|
||||
entry = (tid, track)
|
||||
if track.lost_count == 0 and track.hits >= self.min_hits:
|
||||
measured.append(entry)
|
||||
elif track.lost_count == 0:
|
||||
# hits < min_hits 尚未确认,视为暂态
|
||||
measured.append(entry)
|
||||
else:
|
||||
coasting.append(entry)
|
||||
|
||||
# 优先从实测轨道选
|
||||
chosen = None
|
||||
is_predicted = False
|
||||
|
||||
if measured:
|
||||
# 置信度最高的实测轨道
|
||||
measured.sort(key=lambda kv: kv[1].confidence, reverse=True)
|
||||
chosen = measured[0]
|
||||
elif coasting:
|
||||
# 无实测时选最靠近画面中心的
|
||||
cx_frame = frame_shape[1] / 2.0 if frame_shape else 320.0
|
||||
cy_frame = frame_shape[0] / 2.0 if frame_shape else 240.0
|
||||
coasting.sort(key=lambda kv: (
|
||||
(kv[1].last_center[0] - cx_frame) ** 2 +
|
||||
(kv[1].last_center[1] - cy_frame) ** 2
|
||||
))
|
||||
chosen = coasting[0]
|
||||
is_predicted = True
|
||||
|
||||
if chosen is None:
|
||||
return {
|
||||
"tracked": False,
|
||||
"primary_id": None,
|
||||
"center": None,
|
||||
"box": None,
|
||||
"confidence": 0.0,
|
||||
"is_predicted": False,
|
||||
"lost_frames": 0,
|
||||
"n_active": len(self._tracks),
|
||||
"raw": None,
|
||||
}
|
||||
|
||||
tid, track = chosen
|
||||
cx, cy = track.last_center
|
||||
box = track.last_box
|
||||
conf = track.confidence
|
||||
lost = track.lost_count
|
||||
|
||||
return {
|
||||
"tracked": True,
|
||||
"primary_id": tid,
|
||||
"center": (float(cx), float(cy)),
|
||||
"box": box,
|
||||
"confidence": float(conf),
|
||||
"is_predicted": is_predicted,
|
||||
"lost_frames": lost,
|
||||
"n_active": len(self._tracks),
|
||||
"raw": (None if is_predicted else box),
|
||||
}
|
||||
|
||||
def __repr__(self):
|
||||
tracks_info = ", ".join(str(t) for t in self._tracks.values())
|
||||
return (f"TargetTracker(frame={self._frame_count}, "
|
||||
f"active={len(self._tracks)}, "
|
||||
f"tracks=[{tracks_info}])")
|
||||
+151
-42
@@ -23,6 +23,7 @@ except ImportError:
|
||||
# ================== ZMQ Router 通信代理 =====================
|
||||
from drone_controller.router_proxy import RouterProxy
|
||||
from drone_controller.pid_counter import PID
|
||||
from drone_controller.target_tracker import TargetTracker
|
||||
# ===========================================================
|
||||
|
||||
|
||||
@@ -277,6 +278,17 @@ class UAVControlLoop:
|
||||
self.pid_y = PID(**pid_y_cfg)
|
||||
self.pid_z = PID(**pid_z_cfg)
|
||||
|
||||
# ---- 轻量级目标追踪过滤器 ----
|
||||
# 平滑 YOLO 检测、稳定 ID、coast 预测
|
||||
tracker_cfg = fc_cfg.get("target_tracker", {})
|
||||
self.tracker = TargetTracker(
|
||||
max_lost_frames=tracker_cfg.get("max_lost_frames", 8),
|
||||
max_association_dist=tracker_cfg.get("max_association_dist", 200.0),
|
||||
min_hits=tracker_cfg.get("min_hits", 3),
|
||||
)
|
||||
self._tracked_result = None # 跟踪结果(锁保护)
|
||||
# ------------------------------------------------
|
||||
|
||||
# 控制参数(按协议改为 20Hz,§4.1 要求 ≥20Hz)
|
||||
self.control_hz = 20.0
|
||||
self.control_interval = 1.0 / self.control_hz
|
||||
@@ -285,7 +297,7 @@ class UAVControlLoop:
|
||||
self.running = False
|
||||
self.control_thread = None
|
||||
self.current_target = (0.0, 0.0, 0.0)
|
||||
self.detections = [] # 最新检测结果(线程间共享)
|
||||
self.detections = [] # 最新检测结果(仅用于渲染,线程间共享)
|
||||
self.frame_shape = (480, 640) # 默认尺寸
|
||||
|
||||
# 锁保护共享数据
|
||||
@@ -294,6 +306,9 @@ class UAVControlLoop:
|
||||
# ---- 防断流戳:主线程每次 update_detections 时更新 ----
|
||||
self._last_update_time = 0.0
|
||||
|
||||
# ---- 安全刹车请求标志(主线程判丢 → 控制线程执行刹车) ----
|
||||
self._brake_requested = False
|
||||
|
||||
def start_uav(self):
|
||||
"""连接 Router,执行标准起飞序列(协议 §3.3.1)
|
||||
|
||||
@@ -437,48 +452,95 @@ class UAVControlLoop:
|
||||
def update_detections(self, detections, frame_shape):
|
||||
"""由主循环调用,更新最新检测结果和图像尺寸
|
||||
|
||||
处理流程:
|
||||
1. 用 YOLO 原始检测更新 TargetTracker(获取平滑/预测结果)
|
||||
2. Coast 滑行期间 → PID 积分清空(只靠 PD 稳住,不充积分)
|
||||
3. 判定彻底丢失 → 请求控制线程执行安全零速刹车
|
||||
4. 将结果写入线程安全共享区,供控制线程消费
|
||||
|
||||
Args:
|
||||
detections: postprocess 后的检测列表 [[x1,y1,x2,y2,conf,cls_id], ...]
|
||||
frame_shape: 原始帧形状 (height, width)
|
||||
|
||||
Returns:
|
||||
(tracked: bool, result: dict)
|
||||
— tracked: 当前是否存在有效主目标
|
||||
— result: 跟踪器完整输出(含 center, box, is_predicted, lost_frames 等)
|
||||
"""
|
||||
# ---- 跟踪过滤器:处理原始检测,输出平滑中心点 ----
|
||||
# 计算在锁外进行,避免阻塞控制线程(Kalman 计算 < 0.1ms 量级)
|
||||
result = self.tracker.update(detections, frame_shape)
|
||||
tracked = result.get("tracked", False)
|
||||
# ------------------------------------------------
|
||||
|
||||
# ---- ① Coast 滑行期间:PID 积分清空 ----
|
||||
# 卡尔曼预测阶段目标可能漂移,积分项在此累积会导致
|
||||
# 目标重获瞬间积分暴冲(I-term windup)
|
||||
if tracked and result.get("is_predicted"):
|
||||
self.pid_y.reset_integral()
|
||||
self.pid_z.reset_integral()
|
||||
|
||||
# ---- ② 目标彻底丢失:请求控制线程执行刹车 ----
|
||||
# 注意:不在此处直接调 proxy.send_setpoint 以避免 REQ-REP
|
||||
# 并发冲突(控制线程也在使用同一 REQ socket),改用标志位
|
||||
# 由控制线程在下一 20Hz 周期拾取并发送刹车指令
|
||||
if not tracked:
|
||||
self.pid_y.reset_integral()
|
||||
self.pid_z.reset_integral()
|
||||
with self.data_lock:
|
||||
self._brake_requested = True
|
||||
|
||||
with self.data_lock:
|
||||
self.detections = detections
|
||||
self.frame_shape = frame_shape
|
||||
self._last_update_time = time.time()
|
||||
self._tracked_result = result
|
||||
|
||||
return tracked, result
|
||||
|
||||
# ---- 内部控制线程 ----
|
||||
|
||||
def _control_loop(self):
|
||||
"""独立线程:以 20Hz 发送 SETPOINT,保证 OFFBOARD 保活
|
||||
|
||||
控制流:
|
||||
1. 读取 TargetTracker 的平滑结果(免锁快照)
|
||||
2. 防断流检测(主线程 500ms 无更新 → 强制判丢)
|
||||
3. 跟踪器判丢 → PID 复位 → 零速刹车
|
||||
4. 跟踪器锁定 → 用平滑中心点计算 PID → VELOCITY SETPOINT
|
||||
|
||||
无论是否有目标,每次迭代都发送 VELOCITY SETPOINT。
|
||||
无目标时速度全零(悬停刹车),满足协议 ≥20Hz 要求。
|
||||
"""
|
||||
while self.running:
|
||||
start = time.time()
|
||||
|
||||
# 获取最新的检测结果(加锁)
|
||||
# ---- ① 读取跟踪结果和帧尺寸(加锁快照) ----
|
||||
with self.data_lock:
|
||||
dets = self.detections.copy()
|
||||
tracked = self._tracked_result
|
||||
h, w = self.frame_shape[:2]
|
||||
last_update = self._last_update_time
|
||||
|
||||
# ---- 防断流检测:超过 500ms 无更新则强制零速 ----
|
||||
# 覆盖主线程卡死、摄像头断流等场景
|
||||
# ---- ② 防断流检测 ----
|
||||
# 主线程超过 500ms 未更新 → 强制判丢(覆盖摄像头断流/主线程卡死)
|
||||
if time.time() - last_update > 0.5:
|
||||
if len(dets) > 0:
|
||||
# 只在首次过期时打印,避免刷屏
|
||||
pass # 保留静默刹车
|
||||
dets = [] # 清空检测 → PID 复位 → 零速
|
||||
self.pid_y.reset()
|
||||
self.pid_z.reset()
|
||||
# ------------------------------------------------
|
||||
tracked = None
|
||||
|
||||
# 计算速度指令
|
||||
vx, vy, vz = self._compute_velocity(dets, w, h)
|
||||
# ---- ③ 解析跟踪结果 → 速度指令(使用跟踪器输出的平滑中心点) ----
|
||||
vx, vy, vz = self._velocity_from_tracker(tracked, w, h)
|
||||
|
||||
# 通过 ZMQ 发送 VELOCITY SETPOINT(协议 §3.1 速度控制)
|
||||
# 无论是否有目标,都持续发送以满足 20Hz 保活要求
|
||||
# ---- ③' 安全刹车请求:主线程判丢时紧急抱闸 ----
|
||||
# update_detections 检测到目标彻底丢失后设置此标志,
|
||||
# 控制线程在此拾取并执行一次全零 VELOCITY 刹车
|
||||
with self.data_lock:
|
||||
if self._brake_requested:
|
||||
self._brake_requested = False
|
||||
vx, vy, vz = 0.0, 0.0, 0.0
|
||||
# 同时清空 PID(确保 PD 也归零)
|
||||
self.pid_y.reset()
|
||||
self.pid_z.reset()
|
||||
|
||||
# ---- ④ 通过 ZMQ 发送 VELOCITY SETPOINT(协议 §3.1 速度控制) ----
|
||||
ok, ack = self.proxy.send_setpoint(
|
||||
vx=vx, vy=vy, vz=vz,
|
||||
yaw_rate=0.0,
|
||||
@@ -489,42 +551,49 @@ class UAVControlLoop:
|
||||
# recv 超时或格式错误由 _send_req 内部恢复 REQ socket
|
||||
pass
|
||||
|
||||
# 精确控制循环频率(20Hz → 每 50ms)
|
||||
# ---- ⑤ 精确控制循环频率(20Hz → 每 50ms) ----
|
||||
elapsed = time.time() - start
|
||||
sleep_time = self.control_interval - elapsed
|
||||
if sleep_time > 0:
|
||||
time.sleep(sleep_time)
|
||||
|
||||
def _compute_velocity(self, detections, frame_w, frame_h):
|
||||
"""根据检测结果计算机体速度 (vx, vy, vz)
|
||||
def _velocity_from_tracker(self, tracked, frame_w, frame_h):
|
||||
"""
|
||||
根据跟踪器输出计算机体速度 (vx, vy, vz)
|
||||
|
||||
与原始逻辑相同:
|
||||
- 无目标 → PID 复位 → 返回零速(悬停/刹车)
|
||||
- 有目标 → 取置信度最高的 → 偏差归一化 → PID 输出
|
||||
这是控制线程使用的唯一速度计算路径。
|
||||
平滑后的中心点 → 归一化误差 → PID 输出。
|
||||
|
||||
Args:
|
||||
tracked: tracker.update() 返回的 result dict(或 None)
|
||||
frame_w: 画面宽度(像素)
|
||||
frame_h: 画面高度(像素)
|
||||
|
||||
Returns:
|
||||
(vx, vy, vz): NED 坐标系速度 (m/s)
|
||||
"""
|
||||
if not detections:
|
||||
# 无目标:悬停,并重置 PID 积分
|
||||
if tracked is None or not tracked.get("tracked"):
|
||||
self.pid_y.reset()
|
||||
self.pid_z.reset()
|
||||
return 0.0, 0.0, 0.0
|
||||
|
||||
# 取置信度最高的目标
|
||||
best = max(detections, key=lambda x: x[4])
|
||||
x1, y1, x2, y2, _conf, _cls_id = best
|
||||
target_x = (x1 + x2) / 2
|
||||
target_y = (y1 + y2) / 2
|
||||
img_center_x = frame_w / 2
|
||||
img_center_y = frame_h / 2
|
||||
cx, cy = tracked["center"]
|
||||
img_cx = frame_w / 2.0
|
||||
img_cy = frame_h / 2.0
|
||||
|
||||
err_x = (target_x - img_center_x) / img_center_x
|
||||
err_y = (target_y - img_center_y) / img_center_y
|
||||
err_x = (cx - img_cx) / img_cx
|
||||
err_y = (cy - img_cy) / img_cy
|
||||
|
||||
vy = self.pid_y.update(err_x) # 左右移动
|
||||
vz = self.pid_z.update(err_y) # 上下移动(NED 中正为下)
|
||||
vx = 0.0 # 前后移动
|
||||
vx = 0.0 # 前后维持不变
|
||||
|
||||
# ---- 滑行预测期间:强制清空积分项 ----
|
||||
# 双重保险(update_detections 中也有一道):
|
||||
# 确保 coast 帧只靠 PD 出力,积分项绝不充能
|
||||
if tracked.get("is_predicted"):
|
||||
self.pid_y.reset_integral()
|
||||
self.pid_z.reset_integral()
|
||||
|
||||
return vx, vy, vz
|
||||
|
||||
@@ -543,8 +612,11 @@ class UAVControlLoop:
|
||||
)
|
||||
time.sleep(0.2) # 确保指令发出
|
||||
|
||||
self.tracker.reset()
|
||||
self._tracked_result = None
|
||||
|
||||
self.proxy.close()
|
||||
print("控制已停止,ZMQ 代理已关闭")
|
||||
print("控制已停止,ZMQ 代理已关闭,跟踪器已重置")
|
||||
|
||||
|
||||
# ==================== 主程序入口 ======================
|
||||
@@ -599,26 +671,63 @@ def main():
|
||||
|
||||
# 3. 坐标反求与控制计算
|
||||
detections = detector.postprocess(outputs, ratio, dwdh, orig_shape)
|
||||
controller.update_detections(detections, orig_shape)
|
||||
tracked, track_res = controller.update_detections(detections, orig_shape)
|
||||
|
||||
# 4. 界面渲染
|
||||
frame = detector.draw_boxes(frame, detections)
|
||||
cv2.putText(frame, f"NPU Pure FPS: {fps_pure:.1f}", (20, 30),
|
||||
# ================================================================
|
||||
# 4. "所见即所得" 界面渲染
|
||||
# 控制与渲染的坐标绝对值统一:绘制追踪器的平滑框,不绘制原始 YOLO 框
|
||||
# ┌──────────┬──────────┬──────────┐
|
||||
# │ 状态 │ 框颜色 │ 标签 │
|
||||
# ├──────────┼──────────┼──────────┤
|
||||
# │ TRACK │ 亮绿(0,255,0) │ ID:1 [MEAS] │
|
||||
# │ COAST │ 黄(0,255,255) │ ID:1 [COAST]│
|
||||
# │ LOST │ 不画框 │ 仅文字 │
|
||||
# └──────────┴──────────┴──────────┘
|
||||
# ================================================================
|
||||
|
||||
# 追踪器锁定 → 绘制平滑框
|
||||
if tracked and track_res.get("box") is not None:
|
||||
bx1, by1, bx2, by2 = track_res["box"]
|
||||
is_coast = track_res.get("is_predicted", False)
|
||||
box_color = (0, 255, 255) if is_coast else (0, 255, 0) # COAST=黄, MEAS=绿
|
||||
label = f"ID:{track_res.get('primary_id','?')}"
|
||||
label += " [COAST]" if is_coast else " [MEAS]"
|
||||
|
||||
cv2.rectangle(frame, (int(bx1), int(by1)), (int(bx2), int(by2)),
|
||||
box_color, 3)
|
||||
cv2.putText(frame, label, (int(bx1), int(by1) - 12),
|
||||
cv2.FONT_HERSHEY_SIMPLEX, 0.6, box_color, 2)
|
||||
|
||||
# 左上角状态面板
|
||||
cv2.putText(frame, f"NPU FPS: {fps_pure:.1f}", (20, 30),
|
||||
cv2.FONT_HERSHEY_SIMPLEX, 0.7, (0, 255, 255), 2)
|
||||
if track_res:
|
||||
status_str = ("✅" if tracked else "❌")
|
||||
lost = track_res.get("lost_frames", 0)
|
||||
n_act = track_res.get("n_active", 0)
|
||||
cv2.putText(frame, f"Tracker: {status_str} ID:{track_res.get('primary_id','-')} "
|
||||
f"Lost:{lost}f Act:{n_act}",
|
||||
(20, 55), cv2.FONT_HERSHEY_SIMPLEX, 0.5, (0, 255, 255), 1)
|
||||
|
||||
# 5. 推送至后台异步图传队列
|
||||
streamer.push_to_stream(frame)
|
||||
|
||||
# 6. 周期性打印 Router 状态(每 100 帧)
|
||||
# 6. 周期性打印 Router + Tracker 状态(每 100 帧)
|
||||
if loop_count % 100 == 0:
|
||||
state = controller.proxy.get_latest_state()
|
||||
tracker_info = (f" TK:{'✅' if tracked else '❌'}"
|
||||
f" ID:{track_res.get('primary_id','?')}"
|
||||
f" Lost:{track_res.get('lost_frames',0)}f"
|
||||
f" {'PRED' if track_res.get('is_predicted') else 'MEAS'}"
|
||||
f" Act:{track_res.get('n_active',0)}")
|
||||
if state:
|
||||
d = state["drone"]
|
||||
age_us, _ = controller.proxy.get_state_freshness(state)
|
||||
age_ms = (age_us or 0) / 1000
|
||||
print(f"📡 [Router] mode={d['mode']} armed={d['armed']} "
|
||||
print(f"📡 Router: mode={d['mode']} armed={d['armed']} "
|
||||
f"alt={d['alt_rel']:.1f}m spd={d['ground_speed']:.1f}m/s "
|
||||
f"batt={d['battery']:.0f}% state_age={age_ms:.0f}ms",
|
||||
f"batt={d['battery']:.0f}% age={age_ms:.0f}ms"
|
||||
f"{tracker_info}",
|
||||
end="" if loop_count % 500 != 0 else "\n")
|
||||
|
||||
except KeyboardInterrupt:
|
||||
|
||||
Reference in New Issue
Block a user