Classifying Human Behaviors: Improving Training of Conventional Algorithms

Research and development of human driving behavior prediction is a typical topic for human behavior recognition in cognitive systems, which plays an important role in the development of Advanced Driver Assistance Systems (ADAS). Routines used are based on machine learning algorithms, which have to be trained first. In this contribution, a strategy for improved training is developed to increase the reliability of the training process with respect to the relevant detection and fault alarm parameters. Four algorithms like Support Vector Machines (SVM), Hidden Markov Models (HMM), Artificial Neural Networks (ANN), and Random Forest (RF) are used as examples. Three lane changing behaviors (left/right lane change and lane keeping) are modeled as classifications, and simulated on a highway scene using driving simulator. To improve the prediction performance of the related models, design parameters, which are unknown and need to be set manually before training, are modified. Using the proposed training procedure the most suitable design parameters can be determined automatically to optimize the performance of the algorithms. Based on data achieved from 7 different drivers four improved models are validated. The finally obtained results show that the prediction performance of the four algorithms can be improved using the proposed optimized training procedure.