Auto Exposure on ROI

1git clone https://github.com/luxonis/depthai-python.git
2cd depthai-python/examples
3python3 install_requirements.py

1#!/usr/bin/env python3
2
3from pathlib import Path
4import sys
5import cv2
6import depthai as dai
7import numpy as np
8
9# Press WASD to move a manual ROI window for auto-exposure control.
10# Press N to go back to the region controlled by the NN detections.
11
12# Get argument first
13nnPath = str((Path(__file__).parent / Path('../models/mobilenet-ssd_openvino_2021.4_6shave.blob')).resolve().absolute())
14if len(sys.argv) > 1:
15    nnPath = sys.argv[1]
16
17if not Path(nnPath).exists():
18    import sys
19    raise FileNotFoundError(f'Required file/s not found, please run "{sys.executable} install_requirements.py"')
20
21previewSize = (300, 300)
22
23# Create pipeline
24pipeline = dai.Pipeline()
25
26# Define source and outputs
27camRgb = pipeline.create(dai.node.ColorCamera)
28camRgb.setPreviewSize(*previewSize)
29camRgb.setInterleaved(False)
30
31camControlIn = pipeline.create(dai.node.XLinkIn)
32camControlIn.setStreamName('camControl')
33camControlIn.out.link(camRgb.inputControl)
34
35# Define a neural network that will make predictions based on the source frames
36nn = pipeline.create(dai.node.MobileNetDetectionNetwork)
37nn.setConfidenceThreshold(0.5)
38nn.setBlobPath(nnPath)
39nn.setNumInferenceThreads(2)
40nn.input.setBlocking(False)
41camRgb.preview.link(nn.input)
42
43# Linking
44xoutRgb = pipeline.create(dai.node.XLinkOut)
45xoutRgb.setStreamName("rgb")
46camRgb.preview.link(xoutRgb.input)
47
48nnOut = pipeline.create(dai.node.XLinkOut)
49nnOut.setStreamName("nn")
50nn.out.link(nnOut.input)
51
52# MobilenetSSD label texts
53labelMap = ["background", "aeroplane", "bicycle", "bird", "boat", "bottle", "bus", "car", "cat", "chair", "cow",
54            "diningtable", "dog", "horse", "motorbike", "person", "pottedplant", "sheep", "sofa", "train", "tvmonitor"]
55
56def clamp(num, v0, v1):
57    return max(v0, min(num, v1))
58
59def asControl(roi):
60    camControl = dai.CameraControl()
61    camControl.setAutoExposureRegion(*roi)
62    return camControl
63
64class AutoExposureRegion:
65    step = 10
66    position = (0, 0)
67    size = (100, 100)
68    resolution = camRgb.getResolutionSize()
69    maxDims = previewSize[0], previewSize[1]
70
71    def grow(self, x=0, y=0):
72        self.size = (
73            clamp(x + self.size[0], 1, self.maxDims[0]),
74            clamp(y + self.size[1], 1, self.maxDims[1])
75        )
76
77    def move(self, x=0, y=0):
78        self.position = (
79            clamp(x + self.position[0], 0, self.maxDims[0]),
80            clamp(y + self.position[1], 0, self.maxDims[1])
81        )
82
83    def endPosition(self):
84        return (
85            clamp(self.position[0] + self.size[0], 0, self.maxDims[0]),
86            clamp(self.position[1] + self.size[1], 0, self.maxDims[1]),
87        )
88
89    def toRoi(self):
90        roi = np.array([*self.position, *self.size])
91        # Convert to absolute camera coordinates
92        roi = roi * self.resolution[1] // 300
93        roi[0] += (self.resolution[0] - self.resolution[1]) // 2  # x offset for device crop
94        return roi
95
96    @staticmethod
97    def bboxToRoi(bbox):
98        startX, startY = bbox[:2]
99        width, height = bbox[2] - startX, bbox[3] - startY
100        roi = frameNorm(np.empty(camRgb.getResolutionSize()), (startX, startY, width, height))
101        return roi
102
103# Connect to device and start pipeline
104with dai.Device(pipeline) as device:
105
106    # Output queues will be used to get the rgb frames and nn data from the outputs defined above
107    qControl = device.getInputQueue(name="camControl")
108    qRgb = device.getOutputQueue(name="rgb", maxSize=4, blocking=False)
109    qDet = device.getOutputQueue(name="nn", maxSize=4, blocking=False)
110    frame = None
111    detections = []
112
113    nnRegion = True
114    region = AutoExposureRegion()
115
116    # nn data (bounding box locations) are in <0..1> range - they need to be normalized with frame width/height
117    def frameNorm(frame, bbox):
118        normVals = np.full(len(bbox), frame.shape[0])
119        normVals[::2] = frame.shape[1]
120        return (np.clip(np.array(bbox), 0, 1) * normVals).astype(int)
121
122    def displayFrame(name, frame):
123        for detection in detections:
124            bbox = frameNorm(frame, (detection.xmin, detection.ymin, detection.xmax, detection.ymax))
125            cv2.rectangle(frame, (bbox[0], bbox[1]), (bbox[2], bbox[3]), (255, 0, 0), 2)
126            cv2.putText(frame, labelMap[detection.label], (bbox[0] + 10, bbox[1] + 20), cv2.FONT_HERSHEY_TRIPLEX, 0.5, 255)
127            cv2.putText(frame, f"{int(detection.confidence * 100)}%", (bbox[0] + 10, bbox[1] + 40), cv2.FONT_HERSHEY_TRIPLEX, 0.5, 255)
128        if not nnRegion:
129            cv2.rectangle(frame, region.position, region.endPosition(), (0, 255, 0), 2)
130        cv2.imshow(name, frame)
131
132    while True:
133        # Instead of get (blocking), we use tryGet (non-blocking) which will return the available data or None otherwise
134        inRgb = qRgb.tryGet()
135        inDet = qDet.tryGet()
136
137        if inRgb is not None:
138            frame = inRgb.getCvFrame()
139
140        if inDet is not None:
141            detections = inDet.detections
142
143            if nnRegion and len(detections) > 0:
144                bbox = (detections[0].xmin, detections[0].ymin, detections[0].xmax, detections[0].ymax)
145                qControl.send(asControl(AutoExposureRegion.bboxToRoi(bbox)))
146
147        if frame is not None:
148            displayFrame("rgb", frame)
149
150        key = cv2.waitKey(1)
151        if key == ord('n'):
152            print("AE ROI controlled by NN")
153            nnRegion = True
154        elif key in [ord('w'), ord('a'), ord('s'), ord('d'), ord('+'), ord('-')]:
155            nnRegion = False
156            if key == ord('a'):
157                region.move(x=-region.step)
158            if key == ord('d'):
159                region.move(x=region.step)
160            if key == ord('w'):
161                region.move(y=-region.step)
162            if key == ord('s'):
163                region.move(y=region.step)
164            if key == ord('+'):
165                region.grow(x=10, y=10)
166                region.step = region.step + 1
167            if key == ord('-'):
168                region.grow(x=-10, y=-10)
169                region.step = max(region.step - 1, 1)
170            print(f"Setting static AE ROI: {region.toRoi()} (on frame: {[*region.position, *region.endPosition()]})")
171            qControl.send(asControl(region.toRoi()))
172        elif key == ord('q'):
173            break