|
Identification methods | NN | HMM | Fuzzy control theory | GMM |
|
Algorithm features | To identify various types of driving behaviors located within relatively large amounts of stored time history data. The quality of feature parameters is crucial to the accuracy of NN. | To describe the statistical properties of stochastic processes and to identify inherent invisible states through external observation sequence. | To formulate fuzzy rules based on previous experience and then design model performance in accordance with the expectations of the designer. | A parametric approach to density estimation is able to generate arbitrarily shaped densities. |
|
Model accuracy | Very high | Very high | High | High |
|
Real-time performance | Fair [37] | Very good [37] | Fair [77] | The traditional GMM is poor, and the advanced GMM is good [78]. |
|
Model adaptive | — | — | — | Using the maximum a posteriori (MAP) or Bayesian adaptive algorithm to adjust parameters of GMM, personalized driver behavior model will be obtained. |
|
Disadvantages | There is not a unified feasible method to adjust parameters (e.g., the number of NN layers) but generally subjective adjustments based on the simulation results of the models; training time is long. | HMM is not suitable for long-term forecasting system and requires artificial hypothesis for the sequence distribution of the current states. | Since its fuzzy rules are formulated based on a priori knowledge, the simulation results may deviate from the actual values. | GMM cannot obtain more efficient modeling of the time series of feature vectors than other methods do. |
|
Applications | NN is suitable for pattern recognition that is easy to access to acquire the feature parameters, such as music recognition and speech recognition. | HMM is suitable for pattern recognition with strong time series data, such as driver’s intention recognition and speech recognition. | Fuzzy control theory is suitable for pattern recognition whose parameter range is difficult to determine. | GMM is expert in identifying short-term driving behaviors but not in long-term driving behaviors. If combined with PWARX, the model can have a good performance both in the short- and long-term driving behaviors. |
|