Feedback-based particle filtering
View Patent ↗Methods for estimating a conditional probability distribution for signal states of a non-linear random dynamic process. The filter is based on multiple particles, each defined by a state space model similar to the dynamic process. Each particle is updated on the basis of a control input derived by proportional gain feedback on an innovation process. The innovation process is the difference between an increment in an observed quantity measured by one or more sensors and an average of a function of the particles. The particle filter of the invention may also be applied to filtering problems with data association uncertainty where multiple measurements are obtained, of which at most one originates from a specified target.
1. A method for estimating a conditional probability distribution for signal states of a non-linear random dynamic process associated with a plurality of moving vehicles, the method comprising:
a. tracking the plurality of moving vehicles by receiving sensor measurement data, the sensor measurement data providing a history of observations;
b. initializing a state for each of a plurality of particles, the plurality of particles together comprising an entire population of particles, the states of the plurality of particles collectively probabilistically representing an estimate of a state of the non-linear random dynamic process at an initial time; and
c. updating the state of each of the plurality of particles to obtain a posterior distribution defining a trajectory of each of the plurality of moving vehicles based on a control input derived by proportional gain feedback on an innovation process, wherein the innovation process is the difference between an observed quantity based on sensor data at time t and a quantity in which a function of a particle state is a term.
2. A method in accordance with claim 1 , further comprising acquiring the sensor measurement data from at least one sensor.
3. A method in accordance with claim 2 , wherein the at least one sensor is a remote sensor.
4. A method in accordance with claim 2 , wherein the at least one sensor is a camera.
5. A method in accordance with claim 1 , wherein the control input is a solution to an optimization problem.
6. A method in accordance with claim 1 , wherein the step of updating includes application of proportional feedback based on a gain function.
7. A method in accordance with claim 6 , wherein the gain function is derived by numerical solution of a boundary value problem.
8. A method in accordance with claim 6 , wherein the gain function is derived by numerical solution of an Euler-Lagrange boundary value problem.
9. A method in accordance with claim 6 , wherein the gain function is obtained by analytical approximation.
10. A method in accordance with claim 6 , wherein the gain function is obtained by a Galerkin finite-element algorithm.
11. A method in accordance with claim 1 , wherein the innovation process comprises, in particular, at least a term proportional to an average of predictions due to the entire population of particles.
12. A method in accordance with claim 1 , wherein the sensor measurement data correspond to a plurality of positions of each of the plurality of moving vehicles at least one moving object relative to a fixed frame of reference.
13. A method in accordance with claim 12 , wherein the at least one plurality of moving vehicles are object is a ballistic objects.
14. A method in accordance with claim 12 , wherein none of the at least one plurality of moving vehicles object is a ballistic object.