Detector
The goal of detector is to find turns, line changes, overtakes, obstacle avoidance of vehicle only on the base of mobile phone sensors. The detector should be able to classify those events in real time working on a handset. Most likely the implementation will be a neural network trained with supervised learning. The telemetry of car will be processed on the stage of learning with the above events pointed out manually by a man.
Turn seems to bee the most easy maneuver to detect. The time for which the gyroscope and accelerometer sensors show high values is big enough to easily detect it. Turn detection by means of sensors may be very useful on low speeds or when a car is close to high buildings where gps can give large inaccuracy.
Obstacle avoidance seems to be harder to classify than turns, but still it's a solvable problem. Whenever it occurs a car changes its direction very quickly. Those changes are accompanied with lateral acceleration on the accelerometer sensor and corresponding activity on the gyroscope. Obstacle avoidance may be useful to detect accidents and road potholes. Distributed information about these events can help to automatically set road accidents in navigation software.
Change of line is the hardest to detect but not less useful at that. You can still detect the change because at the start and the end of maneuver the car usually turns slightly stronger. It can be seen on the charts. However in contrast to the previous ones line changes are likely to not be always detectable, since to maneuver may be just very gentle. Line changes may help to guess at what line the car currently is and collect information about total count of lines at this part of road. If we know what line the car is occupying we can calculate traffic jams on separate lines and suggest driver to take other line.
Overtake is one more maneuver that seems to be detectable. Overtake is like obstacle avoidance but more smooth and is usually accompanied with the increase of speed (positive frontal acceleration).
U-turn as a usual turn can easily be detected. It can help strongly to navigator to detect immediately that the direction of movement has changed. Without it, it takes usually some 10 seconds for navigator to find that the direction is now other.
One guess is that it can be done with a neural network that first learns with supervised learning on events from real life and then processes data from sensors in realtime.
For instance it may have n inputs that are mapped to sensors data for some period. The algorithm goes through data to check whether given part of it can be classified as one of the events.