Identify a Vehicle and its Driver from the Vehicle ’ s Maneuver Data

In this project, we applied supervised learning algorithms to detect drivers based on their driving behavior. Using labeled driving data (time, speed, acceleration, heading and gas usage), we created a set of features such as maximum of speed, standard deviation of acceleration as well as additional complex features like variation of heading and variation of MPG over time. In addition, we applied PCA to convert a set of possibly correlated features into a set of values of linearly uncorrelated variables and optimize our algorithm performance. After applying the PCA, the accuracy results were improved from 86.0% to 87.7% and the running time was also decreased (using SVM model). We applied SVM, Random Forest, Decision Tree, Naïve Bayes, and Nearest Neighbor algorithms to the data and compared the accuracy and running time of the results for different features sets. Adding complex features greatly improved the classification results for all algorithms. Among the above mentioned algorithms, SVM optimized for its parameters based on a grid search provided the best accuracy of 96%. However, the running time of SVM was the longest.