#neuromorphic #shared-data #driver #data-processing #camera #usb #devices

neuromorphic-types

Shared Neuromorphic data types for drivers and processing frameworks

4 releases (breaking)

0.4.0 Jul 2, 2023
0.3.0 Jun 26, 2023
0.2.0 Jun 11, 2023
0.1.1 May 7, 2023
0.1.0 May 7, 2023

#772 in Hardware support

Download history 53/week @ 2024-07-29 6/week @ 2024-08-05 78/week @ 2024-08-26 117/week @ 2024-09-02 26/week @ 2024-09-09 20/week @ 2024-09-23 42/week @ 2024-09-30 222/week @ 2024-10-07 201/week @ 2024-10-14 9/week @ 2024-10-21 336/week @ 2024-10-28 118/week @ 2024-11-04

666 downloads per month
Used in neuromorphic-drivers

Custom license

9KB
62 lines

neuromorphic_drivers is a library to interact with USB Neuromorphic devices. The drivers were written from scratch with portability and performance in mind.

Supported devices and features

Name Type Resolution Data types Mask Synchronize Rate limiter
Prophesee EVK4 Camera 1280 × 720 DVS, trigger
Prophesee EVK3 HD Camera 1280 × 720 DVS, trigger -

This table lists fratures supported by this library. Certain sensors support unlisted features or features marked as "no" that have yet to be added to neuromorphic_drivers.

Name Links
Prophesee EVK4 https://www.prophesee.ai/event-camera-evk4/
Prophesee EVK3 HD https://www.prophesee.ai/event-based-evk-3/

Python

Get started

python3 -m pip install neuromorphic_drivers

import neuromorphic_drivers as nd

nd.print_device_list()

with nd.open() as device:
    for status, packet in device:
        # packet = {"dvs_events": np.array([...])}

status contains

Examples

See python/tests for different usage examples. python/tests/display.py implements a live event viewer with exponential decays caculated by the GPU. It requires vispy and glfw (python3 -m pip install vispy glfw).

Performance considerations

Recent event-based cameras, such as the Prophesee EVK4, can generate more data than can be processed in real-time under certain circumstances. While the data can always be moved to the computer's memory in real-time, the simplest algorithms (including converting the raw USB bytes to a {t, x, y, polarity} event representation) struggle to keep up during data rate peaks. This library uses seperate threads for reading (USB to memory) and processing (memory to memory or disk) with a circular buffer (ring) at the interface. Short data bursts are seemlessly absorbed by the ring and are typically not an issue, even though they brifely cause a spike in latency. However, persistent high data rates cause the ring to slowly fill and cause the program to eventually crash. Depending on the use-case, one of the following work arounds can be applied:

  • reduce the camera's sensitivity (usually by tuning diff_off and diff_on)
  • enable the event rate limiter if the sensor supports it (the limiter randomly drops events on the sensor side, reducing bandwidth issues but significantly degrading the quality of transient bursts)
  • reduce the camera's spatial resolution by masking rows and columns
  • change the environment if possible (avoid flickering lights, reduce the optical flow, remove background clutter, keep large and fast objects out of the field of view)
  • call device.clear_backlog(until=0) whenever the backlog becomes too large (the maximum backlog is the size of the ring minus the size of the transfer queue)
  • use nd.open(raw=true) to skip the parser and directly access the USB bytes, typically to save them to a file

(Advanced) Direct Memory Access

This library relies on libusb for all USB communications. libusb supports Direct Memory Access (Linux only for now, though other platforms may be added in the future), which allows the USB controller to directly write packets to memory without requiring CPU (and OS kernel) intervention. While this can increase performance and reduce CPU usage, DMA comes without caveats that we disabled it by default. Users when knoweldge of their USB controller and its limitations may want to enable it to increase performance on embedded systems.

USB drivers have a limited number of DMA file objects (128 / 256 / 512 / 1024). This is typically not enough to accomodate event bursts (we use 4096 buffers by default for the EVK4, with 131072 bytes per buffer). The code fall backs to non-DMA buffers if all the DMA buffers are used (for instance the first 128 buffers would be DMA and the rest would be non-DMA) and may result in variable performance over time.

Using all the available DMA buffers can cause other USB transfers to fail (including control transfers to configure the sensor).

Using DMA thus requires one of the following workarounds:

  • Use a small number of DMA buffers and copy packets to a larger ring before processing (this somewhat defeats the purpose of DMA and increases memory copies).
  • Use larger buffers (around 1 MB) to inccrease the ring size without increasing the number of buffers, at the cost of increased latency.
  • Ensure that processing can always keep up with the data rate (sparse scenes / low-sensitivity biases / event-rate limiter / simple processing). Note that simply expanding vectorized EVT3 events to 13-bytes DVS events is not real-time during data peaks.
cargo publish
cd python
python3 -m venv .venv
source .venv/bin/activate
pip install maturin
maturin develop
cd drivers
cargo test --release read -- --nocapture

/etc/udev/rules.d/65-neuromorphic-drivers.rules

SUBSYSTEM=="usb", ATTRS{idVendor}=="152a",ATTRS{idProduct}=="84[0-1]?", MODE="0666"
SUBSYSTEM=="usb", ATTRS{idVendor}=="04b4",ATTRS{idProduct}=="00f[4-5]", MODE="0666"

sudo udevadm control --reload-rules sudo udevadm trigger

No runtime deps