Skip to Main Content
We consider the problem of automatically developing a 3-dimensional model of the environment based on multiples-dimensional images acquired from differing positions and aspect angles. The problem is complicated by the fact that the sensors are moving, we don't know their precise positions and velocities, and the image fields-of-view may overlap one another in an irregular fashion. This type of problem would be encountered, for example, when attempting to combine information from a swarm of unmanned aerial vehicles to perform automatic target detection, classification, and surveillance. To solve this problem we propose a method whereby a probabilistic model of the preprocessed image data is computed, in which parameters of the model include object locations and classification feature statistics, as well as velocities and positions of the sensors. The parameters are then estimated by maximizing a log-likelihood function which quantitatively measures how well the model fits the data. The crux of the problem is data association, i.e. determining which data samples correspond to which physical objects in the environment. Our approach makes use of a convergent, iterative, system of equations in which data association is performed concurrently with parameter estimation during maximization of the log-likelihood. An advantage of our method is that the computational complexity increases only linearly with the size of the model, and thus the approach is more efficient than the standard approaches.