Mono & MobilenetSSD & Depth

View as Markdown

1#!/usr/bin/env python3
2
3from pathlib import Path
4import sys
5import cv2
6import depthai as dai
7import numpy as np
8
9# Get argument first
10nnPath = str((Path(__file__).parent / Path('../models/mobilenet-ssd_openvino_2021.4_6shave.blob')).resolve().absolute())
11if len(sys.argv) > 1:
12    nnPath = sys.argv[1]
13
14if not Path(nnPath).exists():
15    import sys
16    raise FileNotFoundError(f'Required file/s not found, please run "{sys.executable} install_requirements.py"')
17
18# MobilenetSSD label nnLabels
19labelMap = ["background", "aeroplane", "bicycle", "bird", "boat", "bottle", "bus", "car", "cat", "chair", "cow",
20            "diningtable", "dog", "horse", "motorbike", "person", "pottedplant", "sheep", "sofa", "train", "tvmonitor"]
21
22# Create pipeline
23pipeline = dai.Pipeline()
24
25# Define sources and outputs
26monoRight = pipeline.create(dai.node.MonoCamera)
27monoLeft = pipeline.create(dai.node.MonoCamera)
28stereo = pipeline.create(dai.node.StereoDepth)
29manip = pipeline.create(dai.node.ImageManip)
30nn = pipeline.create(dai.node.MobileNetDetectionNetwork)
31
32nnOut = pipeline.create(dai.node.XLinkOut)
33disparityOut = pipeline.create(dai.node.XLinkOut)
34xoutRight = pipeline.create(dai.node.XLinkOut)
35
36disparityOut.setStreamName("disparity")
37xoutRight.setStreamName("rectifiedRight")
38nnOut.setStreamName("nn")
39
40# Properties
41monoRight.setCamera("right")
42monoRight.setResolution(dai.MonoCameraProperties.SensorResolution.THE_400_P)
43monoLeft.setCamera("left")
44monoLeft.setResolution(dai.MonoCameraProperties.SensorResolution.THE_400_P)
45
46# Produce the depth map (using disparity output as it's easier to visualize depth this way)
47stereo.setDefaultProfilePreset(dai.node.StereoDepth.PresetMode.HIGH_DENSITY)
48stereo.setRectifyEdgeFillColor(0)  # Black, to better see the cutout from rectification (black stripe on the edges)
49# Convert the grayscale frame into the nn-acceptable form
50manip.initialConfig.setResize(300, 300)
51# The NN model expects BGR input. By default ImageManip output type would be same as input (gray in this case)
52manip.initialConfig.setFrameType(dai.ImgFrame.Type.BGR888p)
53
54# Define a neural network that will make predictions based on the source frames
55nn.setConfidenceThreshold(0.5)
56nn.setBlobPath(nnPath)
57nn.setNumInferenceThreads(2)
58nn.input.setBlocking(False)
59
60# Linking
61monoRight.out.link(stereo.right)
62monoLeft.out.link(stereo.left)
63stereo.rectifiedRight.link(manip.inputImage)
64stereo.disparity.link(disparityOut.input)
65manip.out.link(nn.input)
66manip.out.link(xoutRight.input)
67nn.out.link(nnOut.input)
68
69# Connect to device and start pipeline
70with dai.Device(pipeline) as device:
71
72    # Output queues will be used to get the grayscale / depth frames and nn data from the outputs defined above
73    qRight = device.getOutputQueue("rectifiedRight", maxSize=4, blocking=False)
74    qDisparity = device.getOutputQueue("disparity", maxSize=4, blocking=False)
75    qDet = device.getOutputQueue("nn", maxSize=4, blocking=False)
76
77    rightFrame = None
78    disparityFrame = None
79    detections = []
80
81    # nn data, being the bounding box locations, are in <0..1> range - they need to be normalized with frame width/height
82    def frameNorm(frame, bbox):
83        normVals = np.full(len(bbox), frame.shape[0])
84        normVals[::2] = frame.shape[1]
85        return (np.clip(np.array(bbox), 0, 1) * normVals).astype(int)
86
87    # Add bounding boxes and text to the frame and show it to the user
88    def show(name, frame):
89        color = (255, 0, 0)
90        for detection in detections:
91            bbox = frameNorm(frame, (detection.xmin, detection.ymin, detection.xmax, detection.ymax))
92            cv2.putText(frame, labelMap[detection.label], (bbox[0] + 10, bbox[1] + 20), cv2.FONT_HERSHEY_TRIPLEX, 0.5, color)
93            cv2.putText(frame, f"{int(detection.confidence * 100)}%", (bbox[0] + 10, bbox[1] + 40), cv2.FONT_HERSHEY_TRIPLEX, 0.5, color)
94            cv2.rectangle(frame, (bbox[0], bbox[1]), (bbox[2], bbox[3]), color, 2)
95        # Show the frame
96        cv2.imshow(name, frame)
97
98    disparityMultiplier = 255 / stereo.initialConfig.getMaxDisparity()
99
100    while True:
101        # Instead of get (blocking), we use tryGet (non-blocking) which will return the available data or None otherwise
102        if qDet.has():
103            detections = qDet.get().detections
104
105        if qRight.has():
106            rightFrame = qRight.get().getCvFrame()
107
108        if qDisparity.has():
109            # Frame is transformed, normalized, and color map will be applied to highlight the depth info
110            disparityFrame = qDisparity.get().getFrame()
111            disparityFrame = (disparityFrame*disparityMultiplier).astype(np.uint8)
112            # Available color maps: https://docs.opencv.org/3.4/d3/d50/group__imgproc__colormap.html
113            disparityFrame = cv2.applyColorMap(disparityFrame, cv2.COLORMAP_JET)
114            show("disparity", disparityFrame)
115
116        if rightFrame is not None:
117            show("rectified right", rightFrame)
118
119        if cv2.waitKey(1) == ord('q'):
120            break

1#include <iostream>
2
3// Includes common necessary includes for development using depthai library
4#include "depthai/depthai.hpp"
5
6// MobilenetSSD label texts
7static const std::vector<std::string> labelMap = {"background", "aeroplane", "bicycle",     "bird",  "boat",        "bottle", "bus",
8                                                  "car",        "cat",       "chair",       "cow",   "diningtable", "dog",    "horse",
9                                                  "motorbike",  "person",    "pottedplant", "sheep", "sofa",        "train",  "tvmonitor"};
10
11int main(int argc, char** argv) {
12    using namespace std;
13    // Default blob path provided by Hunter private data download
14    // Applicable for easier example usage only
15    std::string nnPath(BLOB_PATH);
16
17    // If path to blob specified, use that
18    if(argc > 1) {
19        nnPath = std::string(argv[1]);
20    }
21
22    // Print which blob we are using
23    printf("Using blob at path: %s\n", nnPath.c_str());
24
25    // Create pipeline
26    dai::Pipeline pipeline;
27
28    // Define sources and outputs
29    auto monoRight = pipeline.create<dai::node::MonoCamera>();
30    auto monoLeft = pipeline.create<dai::node::MonoCamera>();
31    auto stereo = pipeline.create<dai::node::StereoDepth>();
32    auto manip = pipeline.create<dai::node::ImageManip>();
33    auto nn = pipeline.create<dai::node::MobileNetDetectionNetwork>();
34
35    auto disparityOut = pipeline.create<dai::node::XLinkOut>();
36    auto xoutRight = pipeline.create<dai::node::XLinkOut>();
37    auto nnOut = pipeline.create<dai::node::XLinkOut>();
38
39    disparityOut->setStreamName("disparity");
40    xoutRight->setStreamName("rectifiedRight");
41    nnOut->setStreamName("nn");
42
43    // Properties
44    monoRight->setCamera("right");
45    monoRight->setResolution(dai::MonoCameraProperties::SensorResolution::THE_400_P);
46    monoLeft->setCamera("left");
47    monoLeft->setResolution(dai::MonoCameraProperties::SensorResolution::THE_400_P);
48    // Produce the depth map (using disparity output as it's easier to visualize depth this way)
49    stereo->setDefaultProfilePreset(dai::node::StereoDepth::PresetMode::HIGH_DENSITY);
50    stereo->setRectifyEdgeFillColor(0);  // Black, to better see the cutout from rectification (black stripe on the edges)
51    // Convert the grayscale frame into the nn-acceptable form
52    manip->initialConfig.setResize(300, 300);
53    // The NN model expects BGR input. By default ImageManip output type would be same as input (gray in this case)
54    manip->initialConfig.setFrameType(dai::ImgFrame::Type::BGR888p);
55
56    // Define a neural network that will make predictions based on the source frames
57    nn->setConfidenceThreshold(0.5);
58    nn->setBlobPath(nnPath);
59    nn->setNumInferenceThreads(2);
60    nn->input.setBlocking(false);
61
62    // Linking
63    monoRight->out.link(stereo->right);
64    monoLeft->out.link(stereo->left);
65    stereo->rectifiedRight.link(manip->inputImage);
66    stereo->disparity.link(disparityOut->input);
67    manip->out.link(nn->input);
68    manip->out.link(xoutRight->input);
69    nn->out.link(nnOut->input);
70
71    // Connect to device and start pipeline
72    dai::Device device(pipeline);
73
74    // Output queues will be used to get the grayscale / depth frames and nn data from the outputs defined above
75    auto qRight = device.getOutputQueue("rectifiedRight", 4, false);
76    auto qDisparity = device.getOutputQueue("disparity", 4, false);
77    auto qDet = device.getOutputQueue("nn", 4, false);
78
79    cv::Mat rightFrame;
80    cv::Mat disparityFrame;
81    std::vector<dai::ImgDetection> detections;
82
83    // Add bounding boxes and text to the frame and show it to the user
84    auto show = [](std::string name, cv::Mat frame, std::vector<dai::ImgDetection>& detections) {
85        auto color = cv::Scalar(255, 192, 203);
86        // nn data, being the bounding box locations, are in <0..1> range - they need to be normalized with frame width/height
87        for(auto& detection : detections) {
88            int x1 = detection.xmin * frame.cols;
89            int y1 = detection.ymin * frame.rows;
90            int x2 = detection.xmax * frame.cols;
91            int y2 = detection.ymax * frame.rows;
92
93            uint32_t labelIndex = detection.label;
94            std::string labelStr = to_string(labelIndex);
95            if(labelIndex < labelMap.size()) {
96                labelStr = labelMap[labelIndex];
97            }
98            cv::putText(frame, labelStr, cv::Point(x1 + 10, y1 + 20), cv::FONT_HERSHEY_TRIPLEX, 0.5, color);
99            std::stringstream confStr;
100            confStr << std::fixed << std::setprecision(2) << detection.confidence * 100;
101            cv::putText(frame, confStr.str(), cv::Point(x1 + 10, y1 + 40), cv::FONT_HERSHEY_TRIPLEX, 0.5, color);
102            cv::rectangle(frame, cv::Rect(cv::Point(x1, y1), cv::Point(x2, y2)), color, cv::FONT_HERSHEY_SIMPLEX);
103        }
104        // Show the frame
105        cv::imshow(name, frame);
106    };
107
108    float disparityMultiplier = 255 / stereo->initialConfig.getMaxDisparity();
109
110    while(true) {
111        // Instead of get (blocking), we use tryGet (non-blocking) which will return the available data or None otherwise
112        auto inRight = qRight->tryGet<dai::ImgFrame>();
113        auto inDet = qDet->tryGet<dai::ImgDetections>();
114        auto inDisparity = qDisparity->tryGet<dai::ImgFrame>();
115
116        if(inDisparity) {
117            // Frame is transformed, normalized, and color map will be applied to highlight the depth info
118            disparityFrame = inDisparity->getFrame();
119            disparityFrame.convertTo(disparityFrame, CV_8UC1, disparityMultiplier);
120            // Available color maps: https://docs.opencv.org/3.4/d3/d50/group__imgproc__colormap.html
121            cv::applyColorMap(disparityFrame, disparityFrame, cv::COLORMAP_JET);
122            show("disparity", disparityFrame, detections);
123        }
124
125        if(!rightFrame.empty()) {
126            show("rectified right", rightFrame, detections);
127        }
128
129        int key = cv::waitKey(1);
130        if(key == 'q' || key == 'Q') return 0;
131    }
132    return 0;
133}