Luxonis’ $99, Intel Myriad X based “DepthAI” module for robotics is available on CrowdSupply along with DepthAI-based Raspberry Pi HAT, USB adapter, and RPi CM3 B+ equipped boards. DepthAI provides up to 25.5 fps object detection.

Luxonis has gone to Crowd Supply to pitch a neural accelerator module for the Raspberry Pi based on the Movidius Myriad X Vision Processor Unit (VPU). The company claims its DepthAI can offload far more processing from the Raspberry Pi than a Pi mated with Intel’s Myriad X based Intel Neural Compute Stick 2 (NCS2) USB stick accelerator. That’s particularly notable since Intel owns Movidius.

DepthAI, front and back (left) and a DepthAI heatmap with object detection
(click images to enlarge)

In addition to offering the $99 DepthAI module, which can be integrated with any computer, Luxonis offers a $149 RPi HAT Edition equipped with a DepthAI module designed to work with the Raspberry Pi 3, 3B+, or 4. There’s also a $299 RPi Compute Module Edition with a circuit board build around the DepthAI module and Raspberry Pi Compute Module 3B+ and a $149 USB3 Edition board that can plug into any computer.


DepthAI CM3, HAT, and USB3 packages (left to right)
(click image to enlarge)

Options include a 4K@60Hz camera and a pair of 720p@120Hz stereo cameras selling for $74. All shipments are due Feb. 14, 2020. We heard about DepthAI from a CNXSoft report.

NCS2

Developed as an entry in the Microchip Get Launched 2019 design competition for products involving Microchip parts, the DepthAI module can achieve real-time object detection at up to 25.5 fps on a Raspberry Pi 3B+ as opposed to 8.31 fps with an NCS2 on a 3B+, both using Intel’s OpenVINO dev kit and an Intel RealSense D435 camera.

The performance improvement is due to the DepthAI module performing “the heavy lifting of running the AI, generating the depth information and translating it into 3D coordinates, and calculating feature tracking,” says the company. This includes using “Myriad X in the way it was intended to be used — directly attached to cameras over MIPI — thereby unlocking power and capabilities that are otherwise inaccessible.”

DepthAI data path architecture
(click image to enlarge)

DepthAI accepts a neural model from frameworks including TensorFlow and Keras and trained using OpenVINO. It then “runs inference directly on data streaming from the MIPI cameras, and outputs the results of the inference through an easy-to-use Python API.” Unlike the NCS2, DepthAI can process stereo depth, motion estimation, edge detection, Harris filtering, warp/de-warp, JPEG, H.264 and H.265 encoding, and the MIPI ISP pipeline, claims Luxonis.

The DepthAI module can do real-time object localization in addition to detection, which is why it is targeted principally at robotics. As the name suggests, depth perception is also part of the package — while the NCS2 on the Pi 3B+ handles video processing at 30 fps, the DepthAI with Pi can run at 60 fps, thereby enabling stereo depth tracking.

DepthAI Raspberry Pi Compute Module (left) and comparison between DepthAI carriers
(click images to enlarge)

Luxonis does not specifically compare DepthAI to a Myriad X solution running over PCIe instead of USB such as Aaeon’s AI Core X modules. However, it notes: “Existing solutions interface with the Myriad X over USB or PCIE, leaving the MIPI lanes unused and forfeiting the capability to use the most powerful hardware capabilities of the Myriad X.”

The DepthAI module includes power supplies, sequencing, and monitoring components, as well as all the required clock and driver circuitry. There are 2x 2-lane and single 4-lane MIPI-CSI interfaces, as well as UART, USB, GPIO, and other interfaces expressed via a 100-pin connector. Onboard NOR boot flash and EEPROM are optional.

 
DepthAI carrier boards

The pre-integrated DepthAI Raspberry Pi Compute Module board offers a 1.2GHz, quad -A53 RPi CM3 B+ with 1GB of LPDDR2. Since there’s no mention of eMMC, this likely uses the Lite mode that offers microSD support circuitry instead of eMMC.

DepthAI Raspberry Pi Compute Module detail views (legend for both is on right)
(click images to enlarge)

The board builds on the CM3 B+ foundation with the DepthAI module and a pair of pre-calibrated 720p stereo cameras. There’s also a 4K, 60Hz color camera and a Raspberry Pi Camera port.

Real-world ports include Ethernet, 2x USB 2.0, micro-USB “boot,” DisplayPort, and HDMI ports. Other features include a Raspberry Pi 40-pin GPIO header, a 3.5mm audio jack, a 5V input, and a USB 2.0 header. The 16GB preconfigured microSD card boots up running object localization on 20 classes (PASCAL VOC 2012), including people, animals, vehicles, and common objects.

Raspberry Pi HAT Edition (left) and USB3 Edition detail views
(click images to enlarge)

The Raspberry Pi HAT Edition lacks the onboard cameras of the CM3-based board, but offers 3x FFC camera ports and flexible flat cable camera mounting. There’s also a USB 3.0 Type-C port and a 40-pin GPIO header and pass-through.

The USB3 Edition feature list is almost identical to that of the HAT. The board is designed to plug into any computer port via a USB Type-C port. Altium files and other documentation will be provided for all three carrier boards.

 
Further information

The DepthAI module is available for $99 on Crowd Supply for the next 37 days along with carrier board packages based on it including the $149 RPi HAT Edition, $299 RPi Compute Module Edition, and $149 USB3 Edition board. At publication time, it had reached about a quarter of the $55K goal. All shipments are due Feb. 14, 2020. More information may be found on the DepthAI Crowd Supply page and the Luxonis website.