comma ai | Weixing Zhang | A Tour Through Driver Monitoring (DM) | COMMA_CON talks

Name: comma ai | Weixing Zhang | A Tour Through Driver Monitoring (DM) | COMMA_CON talks
Uploaded: 2021-11-06T21:08:28Z
Channel: george hotz archive
Description: Date of video 31 Jul 2021. comma three devkit https://comma.ai/shop/products/three Video title: A Tour Through Driver Monitoring | COMMA_CON Weixing Zha...

george hotz archive · Beginner ·📄 Research Papers Explained ·4y ago

Date of video 31 Jul 2021. comma three devkit https://comma.ai/shop/products/three Video title: A Tour Through Driver Monitoring | COMMA_CON Weixing Zhang social media: - https://linkedin.com/in/weixing-zhang View original video on Youtube: - https://youtu.be/D0JO5dr8cPE Twitter: - https://twitter.com/comma_ai/status/1455990413622398981 comma.ai youtube playlist: - https://youtube.com/playlist?list=PLzFUMGbVxlQth3li9A_lWOT1HSbIHQpE0 Chapters: 00:00 Introduction (A tour through driver monitoring) 00:17 Topics (Introduction, Where we stand, The future) 00:24 Goal of driver monitoring 01:02 Kee…

Watch on YouTube ↗ (saves to browser)

Chapters (39)

Introduction (A tour through driver monitoring)

0:17 Topics (Introduction, Where we stand, The future)

0:24 Goal of driver monitoring

1:02 Keeping the driver engaged

1:30 comma Driver Monitoring: spoken by numbers

1:40 Driver Monitoring vs. no Driver Monitoring 90% decrease in distractions more tha

2:40 engaged vs not engaged (drivers pay 1.6x more attention)

3:39 Keeping the driver engaged summery

3:58 How does everything work

4:02 Driver Monitoring in openpilot

4:17 neural net - driverState

4:37 DriverStatus

4:48 carState, modelData

5:07 BEEPS!

5:21 The model - Efficient B0

5:50 The model - Outputs (head pose, eye states)

6:37 The model training data

7:28 labeling / ground-truthing

7:43 Auto ground-truthing

7:56 Fully automatic data annotation pipeline

8:15 Auto ground-truthing stack in action

8:50 Domain adaptation

9:20 from eon to comma two to comma three

10:12 metrics - Everything not explicitly tracked will regress - Harald

10:23 Real metrics for good UX - accuracy

11:08 Model failure, policy too strict, simply too distracted?

11:42 isDistracted

12:08 Bad lighting, tape over camera, blurry lens?

12:25 visualizing the worst cases

12:51 Worst cases as metrics

12:59 Driver Monitoring as an Machine Learning problem

13:10 Data collection

13:23 Driver Monitoring as a whole

13:54 Solution - Enable Record and Upload Driver Camera

14:14 The future

14:19 End-to-end approach

14:54 Temporal context, 180° big model

15:12 Questions

15:30 How many of you will turn your driver monitoring toggle on?

Playlist

Uploads from george hotz archive · george hotz archive · 0 of 60

← Previous Next →