Processing ROS event camera data in python

This repository holds ROS/ROS2 tools for processing event_camera_msgs under ROS and ROS2 with python. These messages are produced by the metavision_driver and the libcaer_driver. For decoding, the event_camera_codecs package is used internally.

With this repository you can quickly load events from a ROS/ROS2 bag into your python code. The decoder will return a structured numpy array of the same format that the Metavision SDK uses:

dtype={'names':['x','y','p','t'], 'formats':['<u2','<u2','i1','<i4'], 'offsets':[0,2,4,8], 'itemsize':12})]

To access e.g. the timestamps (in microseconds) you would use foo['t'], where foo is the numpy array returned by the decoder. See sample code below.

Supported platforms

Continuous integration is tested under Ubuntu with the following ROS2 distros:

Build Status Build Status Build Status

No package is released for ROS1, but integration is tested under Ubuntu 20.04 and ROS Noetic.

How to install from packages

Under ROS2 you can install the package via

sudo apt-get install ros-${ROS_DISTRO}-event-camera-py

How to build from source

Set the following shell variables:


and follow the instructions here

Decoding event array messages

Here is a sample decoder for ROS2. It uses the BagReader helper class that you can find in the src folder.

from bag_reader_ros2 import BagReader
from event_camera_py import Decoder

topic = '/event_camera/events'
bag = BagReader('foo', topic)
decoder = Decoder()

while bag.has_next():
    topic, msg, t_rec = bag.read_next()
    cd_events = decoder.get_cd_events()
    trig_events = decoder.get_ext_trig_events()

The following sample code shows how to decode event array messages under ROS1.

import rosbag
from event_camera_py import Decoder

topic = '/event_camera/events'
bag = rosbag.Bag('foo.bag')
decoder = Decoder()

for topic, msg, t in bag.read_messages(topics=topic):
    decoder.decode_bytes(msg.encoding, msg.width, msg.height,
	                     msg.time_base, msg.events)
    cd_events = decoder.get_cd_events()
    trig_events = decoder.get_ext_trig_events()

The returned event arrays are structured numpy ndarrays that are compatible with Prophesee’s Metavision SDK.

About timestamps

A message in a recorded rosbag has three sources of time information:

  1. The recording timestamp. This is when the message was written into the bag by the rosbag recorder. It is the least precise of all time stamps and therefore usually not used.

  2. The message time stamp in the header (header.stamp). This is the time when the ROS driver host received the first event packet from the SDK for that ROS message. Remember that a ROS message can contain multiple SDK packets, but the header.stamp refers to the first SDK packet received.

  3. The sensor time encoded in the packets. This time stamp depends on the encoding.

  • For ‘evt3’ (metavision) encoding the raw packet needs to be decoded to obtain the sensor time. The encoded sensor time has two quirks: it wraps around every 2^24 usec (16.77 sec) and it has bit noise errors. The decoder used by the event_camera_py packet keeps track of the wrap around and tries to correct the bit errors. But if you start decoding from the middle of the event stream your sensor time stamps will start at somewhere between 0 and 16.77s due to the wrap around, i.e. sensor time depends on where you start decoding in the message stream.

  • For ‘libcaer_cmp’ (libcaer) encoding, the time stamps in the event stream are in nanoseconds since epoch, which makes them unsuitable for 32 bit representation. For this reason the decoder sets the time stamp of the first event to zero, and all subsequent event times are relative to the first event time. The time since epoch (in usec) of the first event can be obtained from the decoder via get_start_time().

The time ‘t’ column in the python array returned by get_cd_events() is the sensor time 3), in micro seconds. The host time can be obtained by suitably combining the sensor time 3) with the ROS header stamp 2). The most naive way is to compute the time difference between sensor time and header stamp for the first packet and subsequently use that difference to obtain host time from sensor time. Obviously this will not account for drift between sensor and host clocks.

Note that the event time stamps in the structured python array are represented by a 32bit signed integer and thus will roll over after about 35mins!


This software is issued under the Apache License Version 2.0.