1
00:00:00,000 --> 00:00:03,937
Meet Floorsight, our entry for the OpenCV AI Competition.

2
00:00:03,967 --> 00:00:07,378
We help safety teams measure forklift risk from existing cameras,
check the evidence, and

3
00:00:07,408 --> 00:00:10,818
review a warning before alerting a supervisor.

4
00:00:11,480 --> 00:00:14,065
A worker can look almost touching a forklift in a camera image,
while being metres away on

5
00:00:14,095 --> 00:00:16,680
the floor.

6
00:00:16,710 --> 00:00:19,485
Pixel distance can create misleading alarms.

7
00:00:19,515 --> 00:00:22,176
Floorsight maps the scene onto the floor, so a reviewer gets an
estimated distance and its

8
00:00:22,206 --> 00:00:24,866
uncertainty.

9
00:00:25,520 --> 00:00:29,005
OpenCV five handles camera geometry, motion, tracking, risk
zones, privacy filtering and

10
00:00:29,035 --> 00:00:32,520
video.

11
00:00:32,550 --> 00:00:36,412
Separate PyTorch models detect people, forklifts and loads.

12
00:00:36,442 --> 00:00:41,405
We estimate a close pass, including stopping distance, turning
space and hidden areas.

13
00:00:41,435 --> 00:00:44,858
Real footage is processed offline; this is not live braking.

14
00:00:45,520 --> 00:00:49,082
Here is the working dashboard replaying recorded factory footage.

15
00:00:49,112 --> 00:00:53,575
A warning links the camera view, floor map, measurements and
review history.

16
00:00:53,605 --> 00:00:57,567
The agent requests a short replay, a close-up and a second
measurement.

17
00:00:57,597 --> 00:01:01,258
A reviewer can inspect the evidence, then approve or reject the
warning.

18
00:01:01,920 --> 00:01:04,295
The agent does more than describe a picture.

19
00:01:04,325 --> 00:01:07,762
Visual evidence changes what it asks for and recommends next.

20
00:01:07,792 --> 00:01:10,567
A tracking jump alone cannot dismiss a hazard.

21
00:01:10,597 --> 00:01:13,410
Conflicting or missing measurements stay uncertain.

22
00:01:13,440 --> 00:01:16,165
Floor-warning alerts require human approval.

23
00:01:16,195 --> 00:01:21,138
The demonstrated planner is local; live Bedrock access remains
blocked by account settings.

24
00:01:21,800 --> 00:01:25,203
In controlled two-camera scenes, false exit alerts fell from
eight to zero, while all eight

25
00:01:25,233 --> 00:01:28,637
visible hazards still produced alerts.

26
00:01:28,667 --> 00:01:32,380
All sixteen cases without a second camera went to human review.

27
00:01:32,410 --> 00:01:34,935
These are executable synthetic tests with saved evidence and
traces, not a field accuracy

28
00:01:34,965 --> 00:01:37,490
claim.

29
00:01:38,120 --> 00:01:40,874
On eighty synthetic floor warnings, hazardous rejection
recommendations fell from sixteen to

30
00:01:40,904 --> 00:01:43,657
zero.

31
00:01:43,687 --> 00:01:47,150
False candidates recommended for action fell by forty percent.

32
00:01:47,180 --> 00:01:49,417
All cases still require review.

33
00:01:49,447 --> 00:01:51,938
We count uncertainty as actionable, so the score cannot improve
just by hiding difficult

34
00:01:51,968 --> 00:01:54,458
cases.

35
00:01:55,120 --> 00:01:58,332
To test intervention, we built a simulated warehouse.

36
00:01:58,362 --> 00:02:02,500
Without control, the forklift contacts two people who enter its
path.

37
00:02:02,530 --> 00:02:05,359
This lets us test feedback from vision to movement without
claiming that software can stop a

38
00:02:05,389 --> 00:02:08,218
truck in recorded video.

39
00:02:08,880 --> 00:02:13,292
With control on, both in-path contacts were avoided in that saved
scenario.

40
00:02:13,322 --> 00:02:18,147
The cost was slower work: twenty-seven metres travelled instead
of eighty in forty seconds.

41
00:02:18,177 --> 00:02:22,902
The system cannot prevent every contact, or react to a person no
camera sees.

42
00:02:22,932 --> 00:02:25,458
Real forklift integration has not been performed.

43
00:02:26,120 --> 00:02:30,174
The architecture connects cameras, OpenCV perception, floor-risk
estimates and the review

44
00:02:30,204 --> 00:02:34,257
agent on an AWS Graviton instance using COOL.

45
00:02:34,287 --> 00:02:37,997
S3 stores evidence, DynamoDB keeps track history, CloudWatch
records activity, and SNS

46
00:02:38,027 --> 00:02:41,737
delivers approved alerts.

47
00:02:41,767 --> 00:02:45,146
The recorded AWS run exercised all four adapters.

48
00:02:45,800 --> 00:02:48,691
For the COOL track, we publish the imported build, Graviton
configuration and baseline

49
00:02:48,721 --> 00:02:51,612
comparisons.

50
00:02:51,642 --> 00:02:54,609
The historical OpenCV chain improved by one point one times, but
full perception did not

51
00:02:54,639 --> 00:02:57,605
improve.

52
00:02:57,635 --> 00:03:01,822
New quality-checked experiments also found no meaningful pipeline
speedup.

53
00:03:01,852 --> 00:03:04,226
Detector inference remains the main bottleneck.

54
00:03:04,880 --> 00:03:09,771
The public repository includes a self-contained simulator, three
hundred and six passing

55
00:03:09,801 --> 00:03:14,692
tests on both Linux and Windows, evaluation scripts, model
revisions and raw receipts.

56
00:03:14,722 --> 00:03:19,322
Terraform, bootstrap code and a deployment guide are included for
AWS.

57
00:03:19,352 --> 00:03:22,442
No private footage or credentials are needed to run the
simulator.

58
00:03:23,080 --> 00:03:27,442
Floorsight turns existing cameras into measured, reviewable
safety evidence.

59
00:03:27,472 --> 00:03:30,069
Our contribution is the combination of floor geometry, an
evidence-driven agent and human

60
00:03:30,099 --> 00:03:32,697
control.

61
00:03:32,727 --> 00:03:37,002
The code, deployment instructions, demo and limitations are
public.

62
00:03:37,032 --> 00:03:40,882
The next step is independent labelled footage and testing with
real safety teams.