Abstract:

In order to make the steering control level of the intelligent cars as close as possible to the level of human drivers， a nonlinear fitting method for training and learning the trajectory of skilled drivers is proposed. Based on the piecewise polynomial method， the expression models of four typical steering conditions including right turn， U-turn， lane keeping and lane change are constructed， and the adaptive pseudo-spectral method is used to realize effective connection of piecewise trajectories. In order to avoid the problem of the traditional neural network learning algorithm （such as BPNN） that it is necessary to artificially set a large number of network training parameters， and it is easy to produce the local optimal solution， a nonlinear skilled driver's driving trajectory fitting strategy based on improved extreme learning machine （ELM） is proposed. The Kalman filter （KF） algorithm is introduced to filter the ELM output weight matrix， update the stage cycle calculation， optimize the ELM algorithm， and improve the learning accuracy of ELM in multi-collinearity. KFELM， ELM and BPNN are used respectively to perform nonlinear fitting tests on the skilled driver's driving trajectory under different working conditions. The results show that the training precision and test accuracy of KFELM are obviously better than ELM and BPNN， and the learning speed of KFELM is slightly better than that of ELM， and significantly better than that of BPNN. The improved ELM driving model training method provides a theoretical basis for decision control for autonomous vehicles.

Key words: intelligent car, driving trajectory, nonlinear fitting, skilled driver