To assist the driver in proper parking, the parking infrastructure is equipped with parking assistance computer vision model using Sony Spresense board.
The model is developed using Edge Impulse software.
Sony Spresense
Architecture:
- Data Acquisition
- Model Training
- Deployment
Follow the steps mentioned in the below link to update the firmware and bootloader.
https://docs.edgeimpulse.com/docs/sony-spresense
edge-impulse-daemon
once
the board is getting connected. In data acquisition, collect the image of car parking scenario.
Labelling:
Based on front wheel direction, the label will be
- Steer Right
- Steer Left
- Accelerate
The labelling denotes the action that driver needs to take in parking.
Example:
The front wheel alignment is facing towards right side, where the proper parking needs to be steered to "Left" side, so the label of that image will be Steer Left.
Lets say, Steer Right denotes, the front wheel facing towards left side which is wrong parking direction.
Accelerate denotes that the driver is parking in proper way, he/she needs to accelerate to park it.
Steering Perception for Labelling:Once the data acquisition is done with all the three labels ( Steer Right, Steer Left, Accelerate). In create impulse section, configure the below setting.
In feature generation, select the RGB mode.
To find the best fitting Neural network for the collected data, Eon Tuner will make the work simpler.
In Eon Tuner, Set Accuracy as first preference. The Eon tuner will list down the different neural network configurations with resource consumption and accuracy in validation data.
Select the preferred architecture, It will re configure the Run Impulse and NN training section.
When the trained model is tested with new test data, the accuracy will be 81.2%.
The model can be deployed as Sony Spresense firmware, the user can flash the firmware locally.
Comments