First succesfull embeded neural network - comments


It was fun, thanks for Zepan and Wu and Andri for the help.
The project: neural net to detect obstacles.
Purpose: add a good collision avoidance layer to iNav

Maixpy seems solid as I haven’t had a crash.
Performances are good except the first time loading of model is too slow and blocking. Async loading or faster load are necessary to make boot up time acceptable for an embeded system. Note that second load is near instant. More info on that is needed.
The k210 is quite a black box and in certain situations it doesn’t crash but freeze, it would be ok if serial connection was still functional but it is not, so, retrieving meaningful errors is not possible.
Loading a model too large for the k210 freezes the board. It would be great if the LEDs were used to indicate errors if async load and serial connection are not possible. Loop crash is better than a blank freeze for us, developers.
Zepan is working on error handling and I am looking forward to it.

Hardware form factor is good, smaller would be even better for onboard drone. Maybe use the other side of the PCB to halve the lengh of the board, like some esp32 mini boards.
Hardware camera FOV is too narrow and fstop too high so that required swapping with a brighter and wider lens.
Hardware ruggedness is unknown at this point. It will be interesting to see how it handles vibration and voltage fluctuations inherent to drones.

Overall this project is great so I gladly help however I can.


Hi. Andri’s here. Good sharing.
Please kindly share the project code/repo, if possible. Thanks.


Here it is