Threshold based decision-tree for automatic driving maneuver recognition using CAN-Bus signal

There is a growing need to develop automatic signal processing strategies for analysis, labeling, and modeling of massive naturalistic driving data. In order to contribute to the formulation of human centric driving assistive / active safety systems, it is also imperative to understand the human component along with vehicle dynamics. Using massive corpora of on-road real-traffic naturalistic driving data exported from CAN-Bus signals, it is possible to develop maneuver recognition for continuous driving performance evaluation. Although high accuracy of automatic maneuver recognition could be obtained using statistic methods such as Hidden Markov Models and/or Bayesian Statistics, this comes at great costs in computational complexity which can jeopardize the viability of applying such driving assistance in real-time embedded systems. To save computational resources, this study proposes to set up an easier threshold based decision-tree strategy. The method is based on filterbank analysis of time-frequency spectrogram processing of steering wheel angle and vehicle speed signals. It is shown that threshold selection is critical to ensure that specific maneuvers with higher levels of danger receive higher recognition accuracy. Analysis of the proposed maneuver recognition system is performed using data from the UTDrive vehicle corpus, with a 6-way decision-tree maneuver performance ranging from 54-73% depending on the specific maneuver class. It is suggested that lower performing specific maneuvers could subsequently be enhanced with statistical model based methods afterwards for both embedded in-vehicle systems and tagging of massive data sets.