OAK-D Lite AI Camera: Depth + AI on Edge Device Review

The OAK-D Lite (OpenCV AI Kit with Depth) is a remarkable embedded AI camera that combines a colour camera, stereo depth cameras, and an Intel Myriad X Vision Processing Unit (VPU) running at 4 TOPS. This compact device enables real-time depth perception and neural network inference at the edge without requiring a powerful host computer. This review covers OAK-D Lite’s capabilities, DepthAI Python SDK, and practical applications for India robotics and industrial projects.

Table of Contents

• OAK-D Lite Specifications
• DepthAI Python SDK Setup
• Stereo Depth Pipeline
• On-Device Neural Network Inference
• Real-Time Object Detection with Depth
• India Applications
• OAK-D Lite vs Alternatives
• FAQ

OAK-D Lite Specifications

OAK-D Lite hardware:

• Colour camera: IMX214 13MP (4208×3120), 68-degree HFOV
• Stereo cameras: 2x OV7251 1MP (640×480 each), 73-degree HFOV, 75mm baseline
• VPU: Intel Myriad X MA2485, 4 TOPS (16x SHAVE cores)
• Stereo depth range: 19.6cm to ~35m (dependant on disparity settings)
• Interface: USB 3.1 Type-C
• Power: 2.5W typical via USB
• Dimensions: 91.7mm x 27.8mm x 17.5mm, 61g
• India price: Rs 12,000-15,000

DepthAI Python SDK Setup

# Install DepthAI SDK
pip3 install depthai

# Install udev rules for Linux (required)
echo 'SUBSYSTEM=="usb", ATTRS{idVendor}=="03e7", MODE="0666"' | sudo tee /etc/udev/rules.d/80-movidius.rules
sudo udevadm control --reload-rules && sudo udevadm trigger

# Verify device detected
python3 -c "import depthai as dai; print(dai.__version__)"

# Test with built-in demo (shows colour + depth side by side)
python3 -m depthai_demo

Stereo Depth Pipeline

import depthai as dai
import numpy as np
import cv2

# Create pipeline
pipeline = dai.Pipeline()

# Define sources and outputs
camRgb = pipeline.create(dai.node.ColorCamera)
stereo = pipeline.create(dai.node.StereoDepth)
monoLeft = pipeline.create(dai.node.MonoCamera)
monoRight = pipeline.create(dai.node.MonoCamera)

xoutDepth = pipeline.create(dai.node.XLinkOut)
xoutRgb = pipeline.create(dai.node.XLinkOut)
xoutDepth.setStreamName('depth')
xoutRgb.setStreamName('rgb')

# Camera properties
camRgb.setPreviewSize(640, 480)
camRgb.setInterleaved(False)
monoLeft.setCamera('left')
monoLeft.setResolution(dai.MonoCameraProperties.SensorResolution.THE_400_P)
monoRight.setCamera('right')
monoRight.setResolution(dai.MonoCameraProperties.SensorResolution.THE_400_P)

# Stereo depth config
stereo.setDefaultProfilePreset(dai.node.StereoDepth.PresetMode.HIGH_DENSITY)
stereo.initialConfig.setMedianFilter(dai.MedianFilter.KERNEL_7x7)
stereo.setLeftRightCheck(True)
stereo.setSubpixel(False)

# Linking
monoLeft.out.link(stereo.left)
monoRight.out.link(stereo.right)
stereo.depth.link(xoutDepth.input)
camRgb.preview.link(xoutRgb.input)

with dai.Device(pipeline) as device:
    depthQueue = device.getOutputQueue('depth', maxSize=4, blocking=False)
    rgbQueue = device.getOutputQueue('rgb', maxSize=4, blocking=False)
    while True:
        depthFrame = depthQueue.get().getFrame()
        rgbFrame = rgbQueue.get().getCvFrame()
        # Colorise depth for display
        depthVis = cv2.normalize(depthFrame, None, 0, 255, cv2.NORM_MINMAX, cv2.CV_8U)
        depthColor = cv2.applyColorMap(depthVis, cv2.COLORMAP_JET)
        cv2.imshow('Depth', depthColor)
        cv2.imshow('RGB', rgbFrame)
        if cv2.waitKey(1) == ord('q'): break

On-Device Neural Network Inference

OAK-D Lite runs neural networks on its Myriad X VPU. Models must be compiled to OpenVINO Intermediate Representation (IR) format. Luxonis provides pre-compiled models via the DepthAI model zoo:

import blobconverter  # pip3 install blobconverter

# Download pre-compiled YOLOv5n for Myriad X
blob_path = blobconverter.from_zoo(
    name='yolov5n_coco_416x416',
    shaves=6,  # Number of SHAVE cores (6 balanced for OAK-D Lite)
    zoo_type='depthai'
)
print(f'Model cached at: {blob_path}')

Real-Time Object Detection with Depth

import depthai as dai, cv2, numpy as np

pipeline = dai.Pipeline()
detectionNetwork = pipeline.create(dai.node.YoloDetectionNetwork)
camRgb = pipeline.create(dai.node.ColorCamera)
stereo = pipeline.create(dai.node.StereoDepth)
monoLeft = pipeline.create(dai.node.MonoCamera)
monoRight = pipeline.create(dai.node.MonoCamera)
xoutDet = pipeline.create(dai.node.XLinkOut)
xoutDet.setStreamName('detections')

# Configure detection network
detectionNetwork.setBlobPath(blob_path)
detectionNetwork.setConfidenceThreshold(0.5)
detectionNetwork.setNumClasses(80)
detectionNetwork.setCoordinateSize(4)
detectionNetwork.setIouThreshold(0.5)
detectionNetwork.setNumInferenceThreads(2)
detectionNetwork.input.setBlocking(False)

camRgb.setPreviewSize(416, 416)
camRgb.preview.link(detectionNetwork.input)
detectionNetwork.out.link(xoutDet.input)

with dai.Device(pipeline) as device:
    detQ = device.getOutputQueue('detections', 4, False)
    while True:
        dets = detQ.get().detections
        print(f'Detected {len(dets)} objects')
        for d in dets:
            print(f'  Label: {d.label}, Conf: {d.confidence:.2f}, Distance: {d.spatialCoordinates.z/1000:.2f}m')

India Applications

Robotic navigation: Stereo depth enables obstacle avoidance for autonomous robots in warehouses, hospitals, and educational institutions. Several India startups use OAK-D for delivery robot prototypes.

Agricultural inspection: Combine depth data with crop disease detection (visual inspection) to measure plant height and volume alongside disease classification – valuable for India’s vast agricultural sector.

Industrial quality control: 3D point cloud generation from stereo depth enables dimensional inspection of manufactured parts. Much lower cost than laser scanners (Rs 12,000 vs Rs 2-5 lakh).

Accessibility aids: Obstacle detection and distance estimation for visually impaired users. OAK-D Lite’s USB-C connection to a mobile phone enables a wearable navigation aid.

OAK-D Lite vs Alternatives

FAQ

Can OAK-D Lite run YOLO without a host computer?

It can run inference on its VPU autonomously, but needs a USB host to receive results. Connect to a Raspberry Pi Zero 2W (Rs 2,500) for a standalone system with minimal power consumption.

What is the minimum reliable depth range?

19.6cm (about 20cm) for OAK-D Lite. For closer ranges, use structured light cameras or Time-of-Flight sensors. The 75mm baseline limits near-range depth accuracy – at 20cm distance, depth noise is about 5-10mm.

Can I train custom models for OAK-D Lite?

Yes. Train your model with standard frameworks (PyTorch, TensorFlow), export to ONNX, convert to OpenVINO IR, then compile to Myriad X blob format using OpenVINO tools. Total pipeline takes 30-60 minutes for a small model.

Shop Camera & Vision Modules