Check out my jetson nano project using a very small ESP32 board with wlan and camera to stream to the jetson and decode in opencv. Conventional solutions and pipeline building with opencv didn’t work during the development due to empty frames transmitted by the camera, using opencv on the raw stream got it working. For the gesture recognition mediapipe was used, which didn’t require additional training.
The small esp board could be hidden in stuffed animals or the like to build a home controll system without any visible cameras. This way confortably controlling the lights or anything that a raspberry pi pico W can be connected to like magic with your hands is possible.
In the project the camera was hidden in a rather large christsmas ball to fit a larger lithium battery. The system controlled an LED strip on a christmas tree
Check out the links to the descriptions and instructions here:
Link to the github:
Link to youtube