That you could detect that in pocket is kind of mind blowing. I guess minor differences in movement related to operating the vehicle versus sitting passively produces an identifiable signature?
More so, especially if you know the type of vehicle someone owns. Without knowing the type of vehicle, you would need data from both sides to differentiate between passenger and driver. But it should be trivial to detect the gradient (slope) of the sensor curve while the vehicle is turning either direction to determine whether a passenger is on the left or the right side. Doubly so with more than two dimensions of measurements available (vehicles normally sway a bit on one side or the other while turning one direction or the other). Simple physics that every suspension will follow... Your body will also sway a bit in three dimensions differently if the phone is on your person. Using these principles, you could also determine whether someone was siting in the front or the back, too. Knowing the type of vehicle, or given enough variance in measurements (more data points is better), it should be possible to infer the position of the sensor due to the gradient of the arc in the turning radius.
Just spit-balling here... lol
Obviously, it's inference though, it's not 100% accurate, and it doesn't have to be.