mawei/MoreTore
1
0
Fork 0
mirror of https://github.com/MoreTore/openpilot.git synced 2026-08-05 00:05:59 +08:00
Code Issues Packages Projects Releases Wiki Activity

Modeld readme replaced with driving.h

Browse Source
This commit is contained in:
HaraldSchafer
2022-06-02 14:19:19 -07:00
committed by GitHub
parent 90af436a12
commit f07f7cbc8e
1 changed files with 1 additions and 57 deletions
Download Patch File Download Diff File Expand all files Collapse all files
selfdrive/modeld/models/README.md
+1 -57
View File
@@ -24,63 +24,7 @@ To view the architecture of the ONNX networks, you can use [netron](https://netr
### Supercombo output format (Full size: 6472 x float32)
* **plan**
* 5 potential desired plan predictions : 4955 = 5 * 991
* predicted mean and standard deviation of the following values at 33 timesteps : 990 = 2 * 33 * 15
* x,y,z position in current frame (meters)
* x,y,z velocity in local frame (meters/s)
* x,y,z acceleration local frame (meters/(s*s))
* roll, pitch , yaw in current frame (radians)
* roll, pitch , yaw rates in local frame (radians/s)
* probability[^1] of this plan hypothesis being the most likely: 1
* **lanelines**
* 4 lanelines (outer left, left, right, and outer right): 528 = 4 * 132
* predicted mean and standard deviation for the following values at 33 x positions : 132 = 2 * 33 * 2
* y position in current frame (meters)
* z position in current frame (meters)
* **laneline probabilties**
* 2 probabilities[^1] that each of the 4 lanelines exists : 8 = 4 * 2
* deprecated probability
* used probability
* **road-edges**
* 2 road-edges (left and right): 264 = 2 * 132
* predicted mean and standard deviation for the following values at 33 x positions : 132 = 2 * 33 * 2
* y position in current frame (meters)
* z position in current frame (meters)
* **leads**
* 2 hypotheses for potential lead cars : 102 = 2 * 51
* predicted mean and stadard deviation for the following values at 0,2,4,6,8,10s : 48 = 2 * 6 * 4
* x position of lead in current frame (meters)
* y position of lead in current frame (meters)
* speed of lead (meters/s)
* acceleration of lead(meters/(s*s))
* probabilities[^1] this hypothesis is the most likely hypothesis at 0s, 2s or 4s from now : 3
* **lead probabilities**
* probability[^1] that there is a lead car at 0s, 2s, 4s from now : 3 = 1 * 3
* **desire state**
* probability[^1] that the model thinks it is executing each of the 8 potential desire actions : 8
* **meta** [^2]
* Various metadata about the scene : 80 = 1 + 35 + 12 + 32
* Probability[^1] that openpilot is engaged : 1
* Probabilities[^1] of various things happening between now and 2,4,6,8,10s : 35 = 5 * 7
* Disengage of openpilot with gas pedal
* Disengage of openpilot with brake pedal
* Override of openpilot steering
* 3m/(s*s) of deceleration
* 4m/(s*s) of deceleration
* 5m/(s*s) of deceleration
* Probabilities[^1] of left or right blinker being active at 0,2,4,6,8,10s : 12 = 6 * 2
* Probabilities[^1] that each of the 8 desires is being executed at 0,2,4,6s : 32 = 4 * 8
* **pose** [^2]
* predicted mean and standard deviation of current translation and rotation rates : 12 = 2 * 6
* x,y,z velocity in current frame (meters/s)
* roll, pitch , yaw rates in current frame (radians/s)
* **recurrent state**
* The recurrent state vector that is fed back into the GRU for temporal context : 512
[^1]: All probabilities are in logits, so you need to apply sigmoid or softmax functions to get actual probabilities
[^2]: These outputs come directly from the vision blocks, they do not have access to temporal state or the desire input
Read [here](https://github.com/commaai/openpilot/blob/90af436a121164a51da9fa48d093c29f738adf6a/selfdrive/modeld/models/driving.h#L236) for more.
## Driver Monitoring Model