I did a project along these times some time back, training a Hidden Markov Model on accelerometer + gyroscope data sampled every 100ms gave basically 100% classification accuracy. Can share the code with you if you are interested.

I'm having the opposite problem. I don't use a bicycle, but Google Fit thinks I do. I assume it's confusing my driving with biking, somehow.

It's easyish to derive to good accuracy from the accelerometer signal alone (which I believe is always available).