21 December 2017

We are looking for sharp and talented students who have a passion for machine vision and stochastic signal processing. If you are one of them, contact me and we may be able to provide you with a PhD scholarship.

1 March 2017

Check out our conference ICCAIS2017 that will be held in Thailand during Oct 31 - Nov 3, 2017. The papers are indexed by all major indexing databases such as Scopus and IEEExplore.

Supposed that in a multi-object filtering system, the sensor is controllable (e.g. movable). In this context, the stochastic sensor control problem focuses on selecting the best control command from an ensemble of possible commands in such a way that the controlled sensor is expected to return the most informative measurements.

In a multi-target tracking example, assume that there is a sensor that returns detections. The probability of detection and clutter (false alarm) rate and power of measurement noise vary with distance in such a way that a closer target-sensor distance leads to higher probability of detection, less clutter, and more accurate locations for detected target. At any time, the sensor can remain at its current location or move one step in one of several directions. The sensor control solution is expected to make sensor movement decisions that lead the sensor towards the targets. If the targets are pseudo stationary, then the sensor should end up somewhere in the middle of all targets.

What if the sensor cannot get too close to the targets? Then there is a constraint that needs to be considered within the stochastic control solution. We have solved the problem by formulating the constraint as void probabilities and incorporating it in the optimization part of our stochastic control algorithm.

A similar problem arises when we have a netowrk of sensor nodes that are centrally processed with limited communication capacity, and thus only limited number of sensors' measurements being comunicated for processing. An example is shown below, where at each time one sensor should be selected for its measurements to be processed in the update step of a multi-object filter for multi-target tracking. The next sensor node selected in the next time can only be one of the nodes around the current node.

The following example shows a stochastic control solution for multiple movable sensors in a multi-target tracking application. The detection profile (probability of detection, clutter rate and power of noise) are distance dependent, and in each time, the combined motion commands for all sensors should lead to the senors moving towards the targets, so that the fusion of all local posteriors (created after running local Bayesian updates in each multi-object Bayes filter running locally in each sensor node) is most informative about the number of objects (targets) and their states.

Another example of multi-sensor stochastic control in presence of spinning sensors. The deteciton probability and noise power of each sensor depends on the relative angle of the target with respect to the heading angle of the sensor. In this scenario, as the targets move in the scene together, all the sensors are spun by the multi-sensor control algorithm to be pointing towards the targets.

free web counter
Visits since March 1, 2017