Practical driving analytics with smartphone sensors

Sensing various driving behaviors, such as accelerations, brakes, turns, and change lanes — is of great interest to many applications, e.g., understanding drive quality, detecting road conditions, and more. Many such applications rely on using smartphone placed in a vehicle to collect such data for ease of deployment and use. However, several driving analytics techniques in the recent past, including our own, make simplifying assumptions that the smartphone is stably fixed with certain orientation and the car is driving on flat roads. Our deployment experience reveals that existing approaches may cause orientation misalignment and acceleration over/under estimation due to road slopes and human interactions, which lead to significant sensing errors for driving analytics applications. In this paper, we present several innovative techniques to improve the overall accuracy and usability of smartphone sensors. First, we use machine learning techniques to detect smartphone's relative orientation changes caused by human interactions. Second, we design a slope-aware alignment algorithm to improve alignment accuracy. Third, we track the linear acceleration of the vehicle to address acceleration over/under estimation problems. Fourth, we evaluate the tradeoffs between GPS and inertial sensors, and fuse inertial sensors with GPS to improve the overall accuracy and usability. We develop a smartphone application called XSense that adopts the novel techniques to improve the overall accuracy on driving analytics. Our evaluation of XSense is conducted through measurements of more than 2,000 trips (more than 13,000 miles) from 16 drivers in the past three years, and shows that XSense improves the 75-percentile accuracy by 5x comparing with well-tuned inertial sensors in traditional approach.