Path logo

Intelligent Sensor Validation, Sensor Fusion, and Supervisory Control for Tracking Targets and Avoiding Objects in IVHS

People involved in the project:

office:

University of California at Berkeley
Department of Mechanical Engineering
6102 Etcheverry Hall
Berkeley, CA 94720-1740
Tel: (510) 643-8146
email:
aagogino@euler.Berkeley.EDU
alag@pawn.Berkeley.EDU
goebel@ton.Berkeley.EDU

Abstract

We propose a framework for real time monitoring and diagnosis of automated vehicles in IVHS. This supervisory control methodology is concerned with fault detection, fault isolation, and control reconfiguration of the many sensors, actuators and controllers that are used in the control process. The supervisory control architecture operates at two levels of the AVCS: at the regulation and at the coordination level. Supervisory control activities at the regulation layer deal with validation and fusion of the sensory data, and fault diagnosis of the actuators, sensors, and the vehicle itself. Supervisory control activities at the coordination layer deal with detecting potential hazards, recommending the feasibility of potential maneuvers and making recommendations to avert accidents in emergency situations. The supervisory control activities are achieved through five modules, organized in an hierarchical manner.

Due to the inherent noise of the sensor measurements, the latter are uncertain. Even if several sensors are used, it is not straight forward which sensor to trust or to what degree to trust it. If untreated, the quality of the ride suffers (as demonstrated in this animation) and in the worst case leads to unsafe conditions. Two different approaches for validation and fusion have been investigated: 1.) a probabilistic approach, and 2.) a fuzzy approach. The probabilistic approach is Kalman Filter based and assumes Gaussian noise distributions while the fuzzy approach makes no such assumptions. For fusion, the fuzzy approach uses validation gates which allow the assignment of confidence values to each sensor reading, depending on the sensor characteristics particular to each sensor.

Diagnosis of the sensors is performed on-line via fuzzy abductive methods. A ranking scheme calculates fuzzy closeness measures which are used to obtain the most likely failure (if any).

Below you can see validated and fused values for longitudinal distance as well as the original sensor values obtained from radar, sonar, and optical sensors.

* Publications

* Collaboration with other labs

* Our lab

* Want to send us email?

* More about PATH

Check out the platooning movie (© PATH California):

Below you can see a few distance sensors which are mounted on the car

PATH/Kai Goebel /19 January 1996