A sensor fusion filter structure based on RBFNN aided Kalman filter in target positioning
Examensarbete för masterexamen
Embedded electronic system design (MPEES), MSc
Nowadays multiple sensors are mounted in one vehicle to obtain reliable data useful for environment perception, Kalman-filter-based multisensor data fusion is commonly adopted in vehicles for target positioning to provide active safety features to the end user. Kalman filter is the optimal solution to numerous data prediction problems as long as the noise is Gaussian. However, non-Gaussian noise or un-modeled noise contained in filter signals can seriously degrade the filter performance. Without a priori knowledge of noise in the system, tuning the parameters of Kalman filter can be difficult. To improve the accuracy of filter estimates, a sensor fusion system that integrates a Kalman filter with a radial basis function neural network (RBFNN) is presented in the thesis. Both extended Kalman filter (EKF) and converted measurement Kalman filter (CMKF) are implemented to verify the univesality of the system. RBFNN is chosen due to its universal function approximation, simple structure and faster learning speed. An incremental constructive method is adopted to design the RBFNN in the project. Filter status (time interval between measurements and estimation results) as well as host information (steering angle and acceleration) are sent as input of RBFNN. The training target is the difference between the conventional filter output and the ground truth from the DAQ system. After training, the neural network is able to compensate for the estimation error of the Kalman filter. In the project, Normalized Root Mean Square Error (NRMSE), which measures the difference between the filter’s estimates and actual ground truth data, is used as the evaluation criteria for fusion filter performance. After applying RBFNN to the fusion filter system, the NRMSE of the CMKF is reduced by more than 30%. and the NRMSE of the EKF is improved by more than 20%. The promising results proves that a fusion filter combining neural network and conventional Kalman filter can achieve a better performance than the stand-alone conventional Kalman filter. The proposed neural network is a universal tool to compensate for the estimate error of different Kalman filter types. Since the camera was not applied in the test because of hardware failure, some features of sensor fusion could not be verified. For further improvement, plenty of training sets in more complex scenario should be collected for network training to make the fusion filter reliable on the real road. The association algorithm can be updated to achieve multiple object tracking.
Kalman filter , active safety , EKF , CMKF , sensor fusion , radial basis function neural network , target positioning