ROVI-EVS · USER MANUAL

ROVI-EVS User Manual

Set-up, connection, live tracking, recording and playback, CSV output, super slow motion and focusing for the ROVI-EVS evaluation kit, which tracks moving people and vehicles in real time with an event camera.

1. Introduction

ROVI-EVS is an evaluation kit made of an event camera that outputs only the pixels that change (LUCID Triton2 EVS) and FLOYD processing software. It finds moving people and vehicles and shows their ID, box and speed in real time, and saves them to files.

Live tracking view — boxes, IDs, speed (px/s), trails and motion arrows

What an event camera is

A normal camera captures the whole picture at fixed intervals. In an event camera every pixel works on its own and sends a signal (an event) only when its brightness changes. Each event carries a position (x, y), a time in microseconds and whether the pixel got brighter or darker (polarity).

The sensor structure and principle, and how it differs from an RGB camera, are explained with diagrams and video in the Event Camera (EVS) White Paper.

What this manual covers

TaskChapter
Prepare the camera and PC, connect the network2 · 3
Check → live → record → play back4
Read the screen, keyboard controls5
Work with recordings, save results as CSV6 · 7
10,000 fps super slow motion8
Focusing, scene conditions9 · 10
Troubleshooting11

2. Components and specifications

Components

ItemDetails
CameraLUCID Triton2 EVS (TRT009S-EC) · Sony IMX636 · 1280 × 720 · 2.5GigE
Lens6.5 mm C-mount (8 mm also works)
Processing PCx86-64 PC · Ubuntu 22.04 LTS · FLOYD processing software installed
Cable and powerCat6 LAN cable, PoE+ or external 12–24 V supply (see the camera datasheet)

Recommended PC

ItemFile playback onlyLive tracking (recommended)
CPUx86-64, 4 cores, AVX26–8 cores or more, single-core boost of 4.5 GHz or higher
Memory8 GB16 GB
Storage20 GB free on an SSDNVMe SSD, 250 GB or more (for recordings)
NetworkNoneA 2.5GbE port for the camera only
GPUNot neededNot needed (an NVIDIA GPU only for the AI detection option)
MonitorOptional1920 × 1080
Processing speed depends on single-core speed more than on the number of cores. On a slower PC the software lowers its own workload and keeps running instead of stopping (see "DEGRADE" in chapter 11).

Performance

ItemValueBasis
Tracking update rate100 Hz (10 ms)Measured
Tracking processing timep95 3 msMeasured
Live processing limit9.8 M events per second, no lossMeasured
Super slow motion virtual frames10,000 fps (0.1 ms apart)Measured
Dynamic rangeOver 120 dBSensor specification
Pixel latency100 µs or less @ 1000 lxSensor specification
A Windows version is in preparation. For now, use Ubuntu 22.04.

3. Set-up and connection

3.1 Mounting the camera

  1. Fit the lens and fix the camera on a tripod or bracket so that it does not shake. A shaking camera produces events across the whole scene and makes tracking hard.
  2. Choose a distance at which the target looks large enough on screen. Boxes appear for moving targets of roughly 24 px or more.
  3. Focus as described in chapter 9.
LensField of view (H × V)Height covered at 3 m5 m10 m
6.5 mm51° × 30°1.6 m2.7 m5.4 m
8 mm43° × 25°1.3 m2.2 m4.4 m

3.2 Network

  1. Connect the camera directly to a 2.5GbE port on the PC. Avoid sharing the port or switch with other devices.
  2. Turn on jumbo frames (MTU 9000) on that port. Find the port name with ip link.
ip link                                              # find the port the camera is on (e.g. enp3s0)
sudo ip link set dev enp3s0 mtu 9000                 # for this session
nmcli connection modify "<connection name>" 802-3-ethernet.mtu 9000   # keep after reboot
  1. The camera and the PC must be on the same subnet. Check or change the camera IP with IpConfigUtility or ArenaView from the LUCID Arena SDK.
It also ran without errors at MTU 1500 up to 10 M events per second, but 9000 is recommended for headroom.

3.3 Software

~/evs-venv/bin/evs-env check        # check install paths and libraries

To type shorter commands, add the program folder to PATH. The examples in this manual assume this and write evs-demo.

echo 'export PATH="$HOME/evs-venv/bin:$PATH"' >> ~/.bashrc
source ~/.bashrc

4. Quick start

  1. Check — tests the environment, the camera connection, event reception and recording start in turn. Without a camera it stops after the environment checks, which is also normal.
   evs-demo check
  1. Live tracking — a window shows the camera picture and the tracking results. With several cameras, pick one with -s <serial number>.
   evs-demo live
  1. Record — press r in the live window to start or stop recording. To record from the start:
   evs-demo live --record
  1. Play back — view a recording in the same window.
   evs-demo replay runs/recordings/recording_261118_143000.raw

Press q or Esc to close the window.

Where results are saved

LocationContents
runs/<date_time>_<command>/Run log log.txt, summary summary.json, screenshots, CSV results
runs/<date_time>_check/checklist.jsonCheck results
runs/recordings/Recordings recording_<YYMMDD_HHMMSS>.raw

runs/ is created under the folder you run the command in. Add -o <folder> to save elsewhere.

5. Reading the screen

Tracking marks

MarkMeaning
Coloured boxA moving object being tracked
#numberObject ID — the same target keeps the same number
px/sSpeed on screen (pixels per second)
Coloured lineThe path travelled
ArrowsDirection and size of motion (colour = direction, see the colour wheel at bottom right)
Dashed boxA candidate not yet confirmed

Status lines at the top

LineContents
1Mode (LIVE / REPLAY) · file name · replay speed · time t · update rate (Hz) · display fps
2ev events per second (Mev/s = million per second)
3ms time per processing stage and total · backlog time behind · drop amount discarded — 0 is normal
4tracks confirmed · candidate · briefly hidden objects, number of arrows

Views

Number keys change the kind of picture. Choose up to two to see them side by side; press the same key again to remove one.

KeyViewUse
1EventsThe default. Pixels that got brighter or darker, as dots
2CellsGrid cells where events gather
3Filter maskEvents removed as noise
4Time surfaceRecently changed places are brighter — good for seeing the flow of motion
Events view (left) and time surface view (right) side by side

Keyboard

KeyAction
q / EscQuit
SpacePause · resume
rStart · stop recording (live only)
pSave a screenshot (PNG in the run folder)
tShow · hide boxes
iShow · hide IDs
lShow · hide trails
fArrows: grid → per object → off
nShow · hide unconfirmed candidates
hShow · hide the status lines
1 ~ 4Choose views (table above)
< / >Zoom out · in
kLight flicker filter on · off (chapter 10)
[ / ]Halve · double the event rate limit (chapter 10)
xReset tracking

Control window

When live view or playback starts, a control window (Control) opens next to the video window. Clicking a button is the same as pressing the key shown in [ ] under it, and blue buttons are the functions that are on now.

Control window — blue buttons are on. Hatched cells are advanced or development functions not covered in this manual (leave them at their defaults)
Control window — blue buttons are on. Hatched cells are advanced or development functions not covered in this manual (leave them at their defaults)
ButtonKeyWhat it does
RECrStart · stop recording (live only, chapter 6)
PAUSESpacePause · resume
TRACK ALGOTTracking on · off (turning it off resets tracking)
FLOW ALGOFMotion (flow) computation on · off — needed for the arrows
BOXES · ARROWS · TRAILSt · f · lShow boxes · arrows · trails (Tracking marks in chapter 5)
CLUSTERScShow event clusters, the step before tracking
HUDhShow · hide the status lines
VIEW EVENTS · CELLS · TIME1 · 2 · 4Choose views (views table above, up to two side by side)
ACTIVITYaNoise filter that removes isolated points with no neighbours (software)
SW STCsChange the software noise filter mode — each press: off → cut → keep → trail
HW AFKkLight flicker filter (in the camera, chapter 10)
HW STCjNoise filter in the camera on · off
ERC /2 · ERC x2[ · ]Halve · double the event rate limit (chapter 10)
ACC − · ACC +- · =Shorten · lengthen the time events are gathered into one picture (accumulation time) — short for fast objects, long for slow ones
FLOW SLOW · FLOW FAST, · .Switch the flow settings between slow-motion and fast-motion use
ASSOCgSwitch between two ways of linking objects from frame to frame — try it when IDs change often
RESET TRKxReset tracking
SNAPSHOTpSave a screenshot
ZOOM − · ZOOM +< · >Zoom out · in
QUITqQuit
Settings in the control window apply only to the current run. If in doubt, press RESET TRK or start again with the defaults.

6. Recording and playback

Recording

HowCommand
Whenever you like during live viewPress r in the live window (press again to stop)
From the startevs-demo live --record
For a set timeevs-demo record --duration 30 (30 seconds)

Playback

evs-demo replay <file>                           # real-time speed
evs-demo replay <file> --replay-factor 0.5       # half speed (0 = as fast as possible)
evs-demo replay <file> --process-from 2000000 --process-to 5000000   # only 2 s to 5 s (µs)

Playable files: .raw (recordings from this kit), .dat, .hdf5.

7. Saving results (CSV)

Save the tracking results (boxes) as CSV for use in Excel, Python and so on. For fast processing without a window, add --no-display --replay-factor 0.

evs-demo replay <file> --output-bbox-npy boxes.npy --output-bbox-csv boxes.csv
evs-demo replay <file> --output-bbox-npy boxes.npy --output-bbox-csv boxes.csv --no-display --replay-factor 0   # no window, full speed
evs-demo live --output-bbox-npy boxes.npy --output-bbox-csv boxes.csv                                           # save during live view
In this version, --output-bbox-csv writes a CSV only together with --output-bbox-npy (a NumPy file with the same contents is created as well).

A plain file name saves into this run's folder (runs/<date_time>_<command>/).

boxes.csv columns

ColumnMeaningUnit
tTimeµs (from the start of the recording)
x, yTop-left corner of the boxpx
w, hBox width and heightpx
class_idClass number (0 without AI detection)–
track_idObject ID (the #number on screen)–
class_confidenceTracking confidence — rises the longer an object is tracked steadily0–1

Example (--bbox-rate 10):

t,x,y,w,h,class_id,track_id,class_confidence
3070000,374.2,108.8,58.7,18.4,0,1,0.300
3160000,926.8,65.9,23.1,49.9,0,3,0.300

Other outputs

OptionContents
--bbox-rate 10Write boxes 10 times per second (default 50)
--output-sparse-csv-filename flow.csvRaw data behind the motion arrows: x,y,polarity,timestamp,vx,vy,feature_id,center_x,center_y
--output-metrics metrics.jsonlPer-update statistics such as event count and processing time (one JSON per line)
--roi X Y W HProcess only part of the picture

8. Super slow motion

Events are redrawn as frames 0.1 ms apart (a virtual 10,000 fps) and shown next to what a 30 fps camera would have captured. Moments that a normal camera catches in one or two frames — a fast fall, a spin, an impact — can be seen in detail.

Left: as a 30 fps camera would capture it (33 ms exposure) · Right: EVS 10,000 fps frame of the same moment
evs-demo slowmo <file> --start 1.0 --duration 0.3        # 0.3 s from 1.0 s into the file
evs-demo slowmo --live --capture 0.5                     # live: press c in the preview to slow down the last 0.5 s
evs-demo slowmo <file> --start 1.0 --duration 0.3 --out clip.mp4 --stride 10   # save as MP4
Key (player)Action
SpacePause
, / .One frame back · forward
[ / ]Slower · faster
rBack to the start
v30 fps comparison on · off
sSave the current frame as PNG
qClose

9. Focusing

An event camera cannot see a still picture, so you blink a focus pattern on a monitor and focus while the camera watches it.

  1. On another PC or monitor, open the focus pattern page. Choose Siemens star (default) or slanted edge; click the screen to blink it full screen at 10 Hz.
  2. Set the monitor brightness to 100% and display scaling to 100%, and turn off automatic brightness and dynamic contrast.
  3. Place the camera square to the monitor so that the pattern is large in the middle of the picture.
  4. Run the focus tool.
   evs-demo focus --pattern star
  1. A score bar appears. Turn the lens focus ring slowly and lock it where the score is highest.
KeyAction
sSave the current image (PNG) and score (JSON)
rReset the accumulation (after turning the lens a lot)
qQuit
Pattern page address: https://floyd.kr/manual/rovi-evs/focus.html. On a 60 Hz monitor, blink at a divisor of 60 such as 5, 10 or 15 Hz.

10. Scene tips

SituationWhat to do
Dots all over the screen under fluorescent or LED lightsThe camera is seeing the light flicker that eyes cannot. Press k to turn on the light flicker filter (removes the 100 · 120 Hz band).
The PC cannot keep up (backlog and drop rise)Press [ to halve the event rate limit (default 10 M per second). Processing only the needed area with --roi also helps.
Unwanted motion such as leaves or waterProcess only the area of interest with --roi X Y W H.
A target gets no boxIt may be too small (under about 24 px) or too slow. Move closer or change the lens. Press n to see unconfirmed candidates.
The camera rides on a vehicle or robotThe whole background looks like motion. This version assumes a fixed camera.
Because an event camera sees changes in brightness, it catches motion well in dark or backlit scenes. In total darkness, though, there is no change to see.

11. Troubleshooting

SymptomWhat to check
check does not see the camera (devices: (none))Cable and power (camera LED), same subnet (3.2), Arena SDK installed, no other program such as ArenaView using the camera
No window opens (remote session etc.)Run from the desktop, or run with --no-display and save CSV only
DEGRADE at the topThe PC reached its limit and the software lowered its workload. Lower the event rate limit ([), close other programs and check MTU 9000.
drop is not 0As above. Recordings hold the camera's raw events, so you can play them back and process them again later.
The whole picture shimmersLight flicker (k) or camera shake
r does not recordRecording works only in live view (not during playback).
Playback too fast or slowCheck --replay-factor (1 = real time)
The focus score does not riseIs the pattern blinking, is the monitor at 100% brightness, is the pattern large in the middle of the picture?

If a problem continues, contact us with runs/<date_time>_<command>/log.txt and the evs-demo check result (checklist.json) attached.

12. Contact

CompanyFLOYD Inc.
Emailfloyd@floyd.kr
Phone+82-70-8676-7976
Webfloyd.kr — Review request

For evaluation kit loans or application reviews (high-speed falling-object counting, tracking moving objects in backlight or low light, and so on), send us a review request and we will get back to you.