Driving behavior detection and abnormal identification is a cornerstone to assessing drivers and improving driving safety. In this study, we propose a multi-feature agglomerative hierarchical clustering method to detect abnormal driving based on multi-information such as distance, velocity, acceleration, direction, and turning angle. Firstly, several typical abnormal driving behaviors are analyzed and their global and local motion characteristics are extracted, and a feature similarity measurement strategy is proposed based on structural distance. Then, these features are clustered and analyzed according to the Laplacian transform of the feature matrix, which can effectively reduce the dimensionality of clustering data and achieves the automatic acquisition of the number of clusters in a manner of low complexity. Finally, experiment is implemented to test the effectiveness of proposed detection method. The results show that this method can accurately determine five typical abnormal driving behaviors, i.e. overspeed, sudden braking, rapid acceleration/deceleration, and frequent lane change, which demonstrates its potential to improve the abnormal driving behavior and enhance the driving safety.

CICTP