frigate/docker/rockchip/rknn.py

import logging
import numpy as np
import cv2
import cv2.dnn
from hide_warnings import hide_warnings

from frigate.detectors.detection_api import DetectionApi
from frigate.detectors.detector_config import BaseDetectorConfig
from typing import Literal
from pydantic import Field

logger = logging.getLogger(__name__)

DETECTOR_KEY = "rknn"

class RknnDetectorConfig(BaseDetectorConfig):
    type: Literal[DETECTOR_KEY]
    score_thresh: float = Field(default=0.5, ge=0, le= 1, title="Minimal confidence for detection.")
    nms_thresh: float = Field(default=0.45, ge=0, le= 1, title="IoU threshold for non-maximum suppression.")

class Rknn(DetectionApi):
    type_key = DETECTOR_KEY

    def __init__(self, config: RknnDetectorConfig):
        self.height = config.model.height
        self.width = config.model.width
        self.score_thresh = config.score_thresh
        self.nms_thresh = config.nms_thresh

        self.model_path = config.model.path or "/models/yolov8n-320x320.rknn"

        from rknnlite.api import RKNNLite
        self.rknn = RKNNLite(verbose=False)
        if self.rknn.load_rknn(self.model_path) != 0:
            logger.error('Error initializing rknn model.')
        if self.rknn.init_runtime() != 0:
            logger.error('Error initializing rknn runtime.')

    def __del__(self):
        self.rknn.release()

    def postprocess(self, results):
        """
        Processes yolov8 output.

        Args:
        results: array with shape: (1, 84, n, 1) where n depends on yolov8 model size (for 320x320 model n=2100)

        Returns:
        detections: array with shape (20, 6) with 20 rows of (class, confidence, y_min, x_min, y_max, x_max)
        """

        results = np.transpose(results[0,:,:,0]) # array shape (2100, 84)
        classes = np.argmax(results[:, 4:], axis=1) # array shape (2100,); index of class with max confidence of each row
        scores = np.max(results[:, 4:], axis=1) # array shape (2100,); max confidence of each row

        # array shape (2100, 4); bounding box of each row
        boxes = np.transpose(
            np.vstack((
                results[:,0] - 0.5 * results[:,2],
                results[:,1] - 0.5 * results[:,3],
                results[:,2],
                results[:,3])
            )
        )

        # indices of rows with confidence > SCORE_THRESH with Non-maximum Suppression (NMS)
        result_boxes = cv2.dnn.NMSBoxes(boxes, scores, self.score_thresh, self.nms_thresh, 0.5)

        detections = np.zeros((20, 6), np.float32)

        for i in range(len(result_boxes)):
            if i >= 20:
                break

            index = result_boxes[i]
            detections[i] = [
                classes[index],
                scores[index],
                (boxes[index][1]) / self.height,
                (boxes[index][0]) / self.width,
                (boxes[index][1] + boxes[index][3]) / self.height,
                (boxes[index][0] + boxes[index][2]) / self.width
            ]

        return detections

    @hide_warnings
    def inference(self, tensor_input):
        return self.rknn.inference(inputs=tensor_input)

    def detect_raw(self, tensor_input):
        output = self.inference([tensor_input,])
        return self.postprocess(output[0])