move_detect/edge/adapters/hailo_pose.py

"""Hailo yolov8s_pose HEF adapter (실장비 Pi5 + Hailo-8/8L). Implements PosePort.

yolov8s_pose has NO on-chip NMS: the HEF emits 9 raw feature-map tensors (3 strides
x {box DFL (H,W,64), class score (H,W,1, already sigmoid), keypoints (H,W,51)}). This
adapter does the host-side DFL box-decode + keypoint-decode + NMS and remaps to
COCO-17 `list[Person]`.

The pure-numpy decode below is ported from Hailo's MIT-licensed reference
(Hailo-Application-Code-Examples/runtime/python/pose_estimation/pose_estimation_utils.py,
class PoseEstPostProcessing) and locked by tests/test_hailo_pose.py via an injected
fake infer_fn (no hardware needed).

DEVICE VALIDATION (bringup): the HailoRT wiring in `_build_hailo_infer` cannot run off
Pi/hailort. Validate it on device per docs/hardware-bringup.md, and confirm the model's
input (640x640 uint8 RGB, on-chip normalization) and the 9-tensor output shapes against
the actual HEF before trusting inference. The remap/decode math is what the unit tests fix.
"""
import cv2
import numpy as np

from edge.models import Person, BBox, Keypoint

# COCO-17 keypoint order (yolov8_pose): nose, l/r eye, l/r ear, l/r shoulder,
# l/r elbow, l/r wrist, l/r hip, l/r knee, l/r ankle. Emitted in this order, so the
# 17 keypoints map positionally to edge.models.Keypoint without reindexing.
_NUM_KPTS = 17


def _sigmoid(x):
    return 1.0 / (1.0 + np.exp(-x))


def _softmax(x):
    e = np.exp(x - np.max(x, axis=-1, keepdims=True))
    return e / np.sum(e, axis=-1, keepdims=True)


def _letterbox(frame, size):
    """Aspect-preserving resize to size x size with centered padding; BGR->RGB uint8.
    Returns (canvas, scale, (pad_x, pad_y)) so results can be mapped back to `frame`."""
    h, w = frame.shape[:2]
    scale = min(size / w, size / h)
    nw, nh = max(1, round(w * scale)), max(1, round(h * scale))
    resized = cv2.resize(frame, (nw, nh))
    rgb = cv2.cvtColor(resized, cv2.COLOR_BGR2RGB)
    canvas = np.full((size, size, 3), 114, dtype=np.uint8)
    pad_x, pad_y = (size - nw) // 2, (size - nh) // 2
    canvas[pad_y:pad_y + nh, pad_x:pad_x + nw] = rgb
    return canvas, scale, (pad_x, pad_y)


def _pick(outs, size, channels):
    for t in outs:
        if t.shape[1] == size and t.shape[2] == size and t.shape[3] == channels:
            return t
    raise ValueError(f"missing yolov8s_pose tensor for size={size} channels={channels}")


def _iou(box, boxes):
    x1 = np.maximum(box[0], boxes[:, 0]); y1 = np.maximum(box[1], boxes[:, 1])
    x2 = np.minimum(box[2], boxes[:, 2]); y2 = np.minimum(box[3], boxes[:, 3])
    inter = np.clip(x2 - x1, 0, None) * np.clip(y2 - y1, 0, None)
    area = (box[2] - box[0]) * (box[3] - box[1])
    areas = (boxes[:, 2] - boxes[:, 0]) * (boxes[:, 3] - boxes[:, 1])
    union = area + areas - inter
    return np.where(union > 0, inter / union, 0.0)


def _nms(pred, conf_thres, iou_thres, max_det):
    """pred: (N, 4 + 1 + 51) = [xywh, score, 17*(x,y,score)]. Greedy IoU NMS (1 class)."""
    keep_mask = pred[:, 4] > conf_thres
    pred = pred[keep_mask]
    if pred.shape[0] == 0:
        return []
    xywh = pred[:, :4]
    xyxy = np.empty_like(xywh)
    xyxy[:, 0] = xywh[:, 0] - xywh[:, 2] / 2
    xyxy[:, 1] = xywh[:, 1] - xywh[:, 3] / 2
    xyxy[:, 2] = xywh[:, 0] + xywh[:, 2] / 2
    xyxy[:, 3] = xywh[:, 1] + xywh[:, 3] / 2
    scores = pred[:, 4]
    kpts = pred[:, 5:].reshape(-1, _NUM_KPTS, 3)
    order = np.argsort(-scores)
    dets = []
    while order.size > 0 and len(dets) < max_det:
        i = order[0]
        dets.append({"bbox": xyxy[i], "score": float(scores[i]), "kpts": kpts[i]})
        if order.size == 1:
            break
        ious = _iou(xyxy[i], xyxy[order[1:]])
        order = order[1:][ious <= iou_thres]
    return dets


def decode_yolov8_pose(outs, input_size, strides, reg_max,
                       conf_thres, iou_thres, max_det):
    """Host decode of the 9 raw yolov8s_pose tensors -> NMS'd detections (letterbox px).
    Ported from Hailo's MIT PoseEstPostProcessing (decoder + extract + NMS)."""
    reg_range = np.arange(reg_max + 1)
    boxes_all = scores_all = kpts_all = None
    for stride in strides:
        size = input_size // stride
        box = _pick(outs, size, 4 * (reg_max + 1)).astype(np.float32)
        score = _pick(outs, size, 1).astype(np.float32)
        kpt = _pick(outs, size, _NUM_KPTS * 3).astype(np.float32).copy()

        gx, gy = np.meshgrid(np.arange(size) + 0.5, np.arange(size) + 0.5)
        ct_col, ct_row = gx.flatten() * stride, gy.flatten() * stride
        center = np.stack((ct_col, ct_row, ct_col, ct_row), axis=1)  # (HW, 4)

        bd = box.reshape(1, size * size, 4, reg_max + 1)
        bd = _softmax(bd) * reg_range.reshape(1, 1, 1, -1)
        bd = np.sum(bd, axis=-1) * stride                            # (1, HW, 4) distances (px)
        bd = np.concatenate([bd[:, :, :2] * (-1), bd[:, :, 2:]], axis=-1)
        decode_box = center[None] + bd                              # xyxy
        xmin, ymin = decode_box[:, :, 0], decode_box[:, :, 1]
        xmax, ymax = decode_box[:, :, 2], decode_box[:, :, 3]
        xywh = np.transpose(np.array([(xmin + xmax) / 2, (ymin + ymax) / 2,
                                      xmax - xmin, ymax - ymin]), [1, 2, 0])

        sc = score.reshape(1, size * size, 1)
        kp = kpt.reshape(1, size * size, _NUM_KPTS, 3)
        kp[..., :2] *= 2
        kp[..., :2] = stride * (kp[..., :2] - 0.5) + center[:, :2].reshape(size * size, 1, 2)
        kp = kp.reshape(1, size * size, _NUM_KPTS * 3)

        boxes_all = xywh if boxes_all is None else np.concatenate([boxes_all, xywh], axis=1)
        scores_all = sc if scores_all is None else np.concatenate([scores_all, sc], axis=1)
        kpts_all = kp if kpts_all is None else np.concatenate([kpts_all, kp], axis=1)

    pred = np.concatenate([boxes_all, scores_all, kpts_all], axis=2)[0]
    return _nms(pred, conf_thres, iou_thres, max_det)


class HailoPoseAdapter:
    def __init__(self, hef_path, input_size=640, infer_fn=None,
                 score_threshold=0.5, nms_iou=0.7, reg_max=15, strides=(8, 16, 32),
                 min_person_score=0.2, max_detections=100):
        self.hef_path = hef_path
        self.input_size = input_size
        self.score_threshold = score_threshold
        self.nms_iou = nms_iou
        self.reg_max = reg_max
        self.strides = tuple(strides)
        self.min_person_score = min_person_score
        self.max_detections = max_detections
        self._infer_fn = infer_fn if infer_fn is not None else self._build_hailo_infer(hef_path)

    def _build_hailo_infer(self, hef_path):  # pragma: no cover (device-only: needs hailort + HEF)
        """Lazily build the HailoRT inference closure. Validate on device (bringup).
        Returns infer_fn(frame_640_uint8_rgb) -> list of the 9 output tensors."""
        from hailo_platform import HEF, VDevice, FormatType, HailoSchedulingAlgorithm

        params = VDevice.create_params()
        params.scheduling_algorithm = HailoSchedulingAlgorithm.ROUND_ROBIN
        vdevice = VDevice(params)
        hef = HEF(hef_path)
        infer_model = vdevice.create_infer_model(hef_path)
        infer_model.set_batch_size(1)
        infer_model.input().set_format_type(FormatType.UINT8)     # on-chip normalization
        out_infos = hef.get_output_vstream_infos()
        for info in out_infos:
            infer_model.output(info.name).set_format_type(FormatType.FLOAT32)  # dequantized
        configured = infer_model.configure()
        self._hailo_ctx = (vdevice, configured)  # keep alive for the session

        def infer_fn(frame):
            buffers = {info.name: np.empty(infer_model.output(info.name).shape, dtype=np.float32)
                       for info in out_infos}
            bindings = configured.create_bindings(output_buffers=buffers)
            bindings.input().set_buffer(np.ascontiguousarray(frame))
            # Positional, NOT timeout_ms=. Upstream is `def run(self, bindings, timeout)` in
            # every HailoRT that supports Hailo-8 (v4.19.0 / v4.20.0 / v4.23.0 / master), so
            # the keyword form is a TypeError on EVERY frame — and run_loop swallows it
            # (edge/daemon.py:23-24), so it presents as a green unit emitting zero events.
            configured.run([bindings], 10000)
            return [buffers[info.name][None, ...] if buffers[info.name].ndim == 3
                    else buffers[info.name] for info in out_infos]

        return infer_fn

    def infer(self, frame) -> list:
        h, w = frame.shape[:2]
        canvas, scale, (pad_x, pad_y) = _letterbox(frame, self.input_size)
        outs = self._infer_fn(canvas)
        dets = decode_yolov8_pose(outs, self.input_size, self.strides, self.reg_max,
                                  self.score_threshold, self.nms_iou, self.max_detections)
        persons = []
        for d in dets:
            if d["score"] < self.min_person_score:
                continue
            x1, y1, x2, y2 = d["bbox"]
            bbox = BBox((x1 - pad_x) / scale, (y1 - pad_y) / scale,
                        (x2 - pad_x) / scale, (y2 - pad_y) / scale)
            keypoints = []
            for j in range(_NUM_KPTS):
                kx, ky, ksc = d["kpts"][j]
                keypoints.append(Keypoint(float((kx - pad_x) / scale),
                                          float((ky - pad_y) / scale),
                                          float(_sigmoid(ksc))))
            persons.append(Person(bbox, keypoints))
        return persons