Unknown

#!/usr/bin/env python3
"""
Mono Camera Object Detection using DepthAI YOLO
Reads configuration from JSON and runs inference.
"""

import cv2
import depthai as dai
import numpy as np
import json
import time
import os
from pathlib import Path
from datetime import datetime
import blobconverter

def run_inference(model_path, config_path):
    # Load configuration file
    configPath = Path(config_path)
    if not configPath.exists():
        raise ValueError(f"Path {configPath} does not exist!")

    with configPath.open() as f:
        config = json.load(f)
    
    nnConfig = config.get("nn_config", {})

    # Parse input size
    if "input_size" in nnConfig:
        W, H = tuple(map(int, nnConfig.get("input_size").split('x')))

    # Extract YOLO-specific metadata
    metadata = nnConfig.get("NN_specific_metadata", {})
    classes = metadata.get("classes", 80)  # Default to 80 if missing
    coordinates = metadata.get("coordinates", 4)
    anchors = metadata.get("anchors", [])
    anchorMasks = metadata.get("anchor_masks", {})
    iouThreshold = metadata.get("iou_threshold", 0.5)
    confidenceThreshold = metadata.get("confidence_threshold", 0.5)

    # Parse labels
    nnMappings = config.get("mappings", {})
    labels = nnMappings.get("labels", {})

    # Check if model exists, else download from DepthAI zoo
    nnPath = model_path
    if not Path(nnPath).exists():
        print(f"No blob found at {nnPath}. Downloading from DepthAI Model Zoo...")
        nnPath = str(blobconverter.from_zoo(model_path, shaves=6, zoo_type="depthai", use_cache=True))

    # Create DepthAI Pipeline
    pipeline = dai.Pipeline()

    # Define Mono Camera
    mono = pipeline.create(dai.node.MonoCamera)
    mono.setResolution(dai.MonoCameraProperties.SensorResolution.THE_720_P)
    mono.setFps(30)

    # Image Manipulation to Resize Input to NN
    manip = pipeline.create(dai.node.ImageManip)
    manip.initialConfig.setResize(W, H)
    manip.initialConfig.setFrameType(dai.ImgFrame.Type.BGR888p)
    manip.setMaxOutputFrameSize(W*H*3)  # Convert mono to RGB

    mono.out.link(manip.inputImage)

    # Define YOLO Detection Network
    detectionNetwork = pipeline.create(dai.node.YoloDetectionNetwork)
    detectionNetwork.setBlobPath(nnPath)
    detectionNetwork.setConfidenceThreshold(confidenceThreshold)
    detectionNetwork.setNumClasses(classes)
    detectionNetwork.setCoordinateSize(coordinates)
    detectionNetwork.setAnchors(anchors)
    detectionNetwork.setAnchorMasks(anchorMasks)
    detectionNetwork.setIouThreshold(iouThreshold)
    detectionNetwork.setNumInferenceThreads(2)
    detectionNetwork.input.setBlocking(False)

    # Linking Mono Camera to NN
    manip.out.link(detectionNetwork.input)

    # Define Output Streams
    xoutMono = pipeline.create(dai.node.XLinkOut)
    xoutNN = pipeline.create(dai.node.XLinkOut)
    xoutMono.setStreamName("mono")
    xoutNN.setStreamName("nn")

    # Linking Outputs
    detectionNetwork.passthrough.link(xoutMono.input)
    detectionNetwork.out.link(xoutNN.input)

    # Create output directory with timestamp
    timestamp = datetime.now().strftime("%Y%m%d_%H%M%S")
    output_dir = "output_mono"
    os.makedirs(output_dir, exist_ok=True)

    raw_video_path = os.path.join(output_dir, f"raw_{timestamp}.avi")
    detected_video_path = os.path.join(output_dir, f"detected_{timestamp}.avi")

    # Define Video Writers
    fourcc = cv2.VideoWriter_fourcc(*'XVID')
    fps = 30
    raw_writer = cv2.VideoWriter(raw_video_path, fourcc, fps, (W, H))
    detected_writer = cv2.VideoWriter(detected_video_path, fourcc, fps, (W, H))

    # Helper function for normalization
    def frameNorm(frame, bbox):
        normVals = np.full(len(bbox), frame.shape[0])
        normVals[::2] = frame.shape[1]
        return (np.clip(np.array(bbox), 0, 1) * normVals).astype(int)

    def displayFrame(name, frame, detections):
        color = (255, 0, 0)
        for detection in detections:
            bbox = frameNorm(frame, (detection.xmin, detection.ymin, detection.xmax, detection.ymax))
            cv2.rectangle(frame, (bbox[0], bbox[1]), (bbox[2], bbox[3]), color, 2)
            cv2.putText(frame, labels[detection.label], (bbox[0] + 10, bbox[1] + 20),
                        cv2.FONT_HERSHEY_TRIPLEX, 0.5, 255)
            cv2.putText(frame, f"{int(detection.confidence * 100)}%", (bbox[0] + 10, bbox[1] + 40),
                        cv2.FONT_HERSHEY_TRIPLEX, 0.5, 255)

        detected_writer.write(frame)
        cv2.imshow(name, frame)

    # Start Pipeline Execution
    with dai.Device(pipeline) as device:



        ir_emitter_brightness = 1200  # Set between 0 and 1500 (max)
        device.setIrLaserDotProjectorBrightness(ir_emitter_brightness)
        device.setIrFloodLightBrightness(ir_emitter_brightness)



        qMono = device.getOutputQueue(name="mono", maxSize=4, blocking=False)
        qDet = device.getOutputQueue(name="nn", maxSize=4, blocking=False)

        frame = None
        detections = []
        startTime = time.monotonic()
        counter = 0

        while True:
            inMono = qMono.get()
            inDet = qDet.get()

            if inMono is not None:
                frame = inMono.getCvFrame()
                raw_writer.write(frame)

                cv2.putText(frame, "NN FPS: {:.2f}".format(counter / (time.monotonic() - startTime)),
                            (2, frame.shape[0] - 4), cv2.FONT_HERSHEY_TRIPLEX, 0.4, (255, 255, 255))

            if inDet is not None:
                detections = inDet.detections
                counter += 1

            if frame is not None:
                displayFrame("Mono Camera Detection", frame, detections)

            if cv2.waitKey(1) == ord('q'):
                break

    # Release Resources
    raw_writer.release()
    detected_writer.release()
    cv2.destroyAllWindows()



model_path = './model/drone_rc_birds_1000epoch/best_dorne_rc_birds_1000_openvino_2022.1_6shave.blob'
config_path = './model/drone_rc_birds_1000epoch/best_dorne_rc_birds_1000.json'
# Example usage
# model_path = './model/blob_640_nano_rc_b_d_1k/best_openvino_2022.1_6shave.blob'
# config_path = './model/blob_640_nano_rc_b_d_1k/best.json'
run_inference(model_path, config_path)

hi @jakaskerl whether the code given below for detection using the mono camera is correct?
is it activating the ir flood light to detect in the night?

thanks @jakaskerl , Actually my aim is to use color camera and process the detections on the grayscale frame, the model i am using is trained on 3 channel images.

hi @jakaskerl ,Thanks for the support.
I want to use the model trained with the 3 channels images, can we convert the grayscale frame to 3-channel before passing to the BLOB model for detection, if yes , how?

Can you please give me correct code with all the fixes done.

hi @jakaskerl , even on doing "Instead of camRgb.setPreviewSize(W, H), do camRgb.setVideoSize(W, H)".

I m getting error for the above fix - [1844301051C738F500] [1.3] [4.342] [DetectionNetwork(2)] [error] Input tensor 'images' (0) exceeds available data range. Data size (802816B), tensor offset (0), size (2408448B) - skipping inference

@jakaskerl getting this error continuously…
[1844301051C738F500] [1.3] [3.097] [ImageManip(1)] [error] Output image is bigger (2073600B) than maximum frame size specified in properties (802816B) - skipping frame. Please use the setMaxOutputFrameSize API to explicitly config the [maximum] output size.

please note: the model is trained on 896*896 resolution.

please give the code which can run such models without any errors.

#!/usr/bin/env python3
"""
The code is edited from docs (https://docs.luxonis.com/projects/api/en/latest/samples/Yolo/tiny_yolo/)
We add parsing from JSON files that contain configuration and process detections on a grayscale frame.
"""

from pathlib import Path
import cv2
import depthai as dai
import numpy as np
import time
import json
import blobconverter
import os

def run_inference(model_path, config_path):
    # parse config
    configPath = Path(config_path)
    if not configPath.exists():
        raise ValueError("Path {} does not exist!".format(configPath))

    with configPath.open() as f:
        config = json.load(f)
    nnConfig = config.get("nn_config", {})

    # parse input shape
    if "input_size" in nnConfig:
        W, H = tuple(map(int, nnConfig.get("input_size").split('x')))

    # extract metadata
    metadata = nnConfig.get("NN_specific_metadata", {})
    classes = metadata.get("classes", {})
    coordinates = metadata.get("coordinates", {})
    anchors = metadata.get("anchors", {})
    anchorMasks = metadata.get("anchor_masks", {})
    iouThreshold = metadata.get("iou_threshold", {})
    confidenceThreshold = metadata.get("confidence_threshold", {})

    # parse labels
    nnMappings = config.get("mappings", {})
    labels = nnMappings.get("labels", {})

    # get model path
    nnPath = model_path
    if not Path(nnPath).exists():
        print("No blob found at {}. Looking into DepthAI model zoo.".format(nnPath))
        nnPath = str(blobconverter.from_zoo(model_path, shaves=6, zoo_type="depthai", use_cache=True))

    # Create pipeline
    pipeline = dai.Pipeline()

    # Define sources and outputs
    camRgb = pipeline.create(dai.node.ColorCamera)
    manip = pipeline.create(dai.node.ImageManip)
    detectionNetwork = pipeline.create(dai.node.YoloDetectionNetwork)
    xoutRgb = pipeline.create(dai.node.XLinkOut)
    nnOut = pipeline.create(dai.node.XLinkOut)

    xoutRgb.setStreamName("rgb")
    nnOut.setStreamName("nn")

    # Camera properties
    camRgb.setPreviewSize(W, H)
    camRgb.setResolution(dai.ColorCameraProperties.SensorResolution.THE_1080_P)
    camRgb.setInterleaved(False)
    camRgb.setColorOrder(dai.ColorCameraProperties.ColorOrder.RGB)
    camRgb.setFps(30)

    # Convert to grayscale
    manip.initialConfig.setFrameType(dai.ImgFrame.Type.GRAY8)
    manip.setMaxOutputFrameSize(W * H )  # Ensure correct output size allocation for grayscale
    camRgb.video.link(manip.inputImage)

    # Network specific settings
    detectionNetwork.setConfidenceThreshold(confidenceThreshold)
    detectionNetwork.setNumClasses(classes)
    detectionNetwork.setCoordinateSize(coordinates)
    detectionNetwork.setAnchors(anchors)
    detectionNetwork.setAnchorMasks(anchorMasks)
    detectionNetwork.setIouThreshold(iouThreshold)
    detectionNetwork.setBlobPath(nnPath)
    detectionNetwork.setNumInferenceThreads(2)
    detectionNetwork.input.setBlocking(False)

    # Linking grayscale image to the detection network
    manip.out.link(detectionNetwork.input)
    detectionNetwork.passthrough.link(xoutRgb.input)
    detectionNetwork.out.link(nnOut.input)

    # Connect to device and start pipeline
    with dai.Device(pipeline) as device:
        qRgb = device.getOutputQueue(name="rgb", maxSize=4, blocking=False)
        qDet = device.getOutputQueue(name="nn", maxSize=4, blocking=False)

        frame = None
        detections = []
        startTime = time.monotonic()
        counter = 0

        while True:
            inRgb = qRgb.get()
            inDet = qDet.get()

            if inRgb is not None:
                frame = inRgb.getCvFrame()
                cv2.putText(frame, "NN fps: {:.2f}".format(counter / (time.monotonic() - startTime)),
                            (2, frame.shape[0] - 4), cv2.FONT_HERSHEY_TRIPLEX, 0.4, (255, 255, 255))
                print(counter / (time.monotonic() - startTime))
            
            if inDet is not None:
                detections = inDet.detections
                counter += 1

            if frame is not None:
                cv2.imshow("Grayscale Detection", frame)

            if cv2.waitKey(1) == ord('q'):
                break

# Example usage
model_path = '../model/896_yolov5n6_blob/best_openvino_2022.1_6shave.blob'
config_path = '../model/896_yolov5n6_blob/best.json'
run_inference(model_path, config_path)

hi @jakaskerl
the above code is not working if the model is trained on img size 896*896.

hi @erik can we pass grayscale frame to the blob yolov5 model for detection using the color camera? if yes, how?

Note: We want blob model to process the detections on the grayscale frame instead of the rgb frame using the color camera.

hi @erik can we pass grayscale frame to the blob yolov5 model for detection using the color camera? if yes, how?

Note: We want blob model to process the detections on the grayscale frame instead of the rgb frame using the color camera.

i am not able to detect oakd camera on jetson orin nano with jetpack 6.1, can you please give me sample code with dependencies required.

Sounds good! @JanCuhel

Thanks.

Hi @jakaskerl its a .pt file only. Please click on the black pop out icon above "file size", a new tab will open then you will be able to download the .pt file.

Thanks.

Hey @jakaskerl here is the link to download the model. it will be great if custom model trained using yolov6n6 can also be converted to blob.

hey @jakaskerl , thankyou.

i face error while converting yolovn6 pretrained to blob. the error is "Error while converting to openvino". i am using tools.luxonis.com

Hey @erik
I have trained a custom model using YOLOv5n. While converting the model to a blob format using RVC2, I encountered an error. However, when I used RVC3 for the conversion, it successfully generated the blob. Unfortunately, when running this blob on DepthAI, I received the following error:

"RuntimeError: BlobReader error: File does not seem to be a supported neural network blob."

It's worth noting that a pre-trained YOLOv5n model was converted using RVC2 without issues and ran successfully as a blob. The issue only arises with the custom-trained model.

Could you please assist in troubleshooting this problem?

@erik Thank you.

hi @erik, Could implementing a frame-skipping mechanism, where the system skips a number of frames before processing a frame using the blob model for detection, help in increasing the effective frames per second (FPS) of the object detection process?

@erik for the above question, the custom models are trained on yolov5s.pt