3D forward sensor modeling and application to occupancy grid based sensor fusion

A Generative Model for Online Depth Fusion

“…We compute each depth independently, casting the ray along which the depth is required into the occupancy grid, and outputting the distance to the maximum in the range [1, N ] of the posterior visibility distribution, i.e. equation (12), but using parameters γ i etc. The position of each depth output is refined by fitting a quadratic function to the mode.…”

“…A recent evaluation [9] of these approaches highlights Konolige's [8], which incorporates some reasoning on visibility, but based on measurements rather than current occupancies. Forward sensor models [10][11][12][13] are generative models of the sensor measurement process, which naturally incorporate visibility constraints, but all previous such models have deficiencies. Thrun [13] uses the iterative EM algorithm to update the latent correspondence of measurement to variable, making the method slow.…”

“…Pathak et al . [12] present a closed-form update to probabilities of visibility, and assume that these directly determine occupancy probabilities. As a result, their model has no notion of occluded surfaces; the occupancy updates include an ad-hoc rule to overcome this.…”

“…Many are able to use any given measurement model, e.g. [12,13], such as those learned from ground truth data; [5] uses a Gaussian plus uniform model. The limitation of these methods is that the exact measurement model must be given prior to inference.…”

“…Whilst very similar to that of Pathak et al . [12], it deals probabilistically with occluded surfaces, and also allows us to associate a different and unknown measurement model per voxel, similar to [17]. These unknown measurement models are inferred in parallel with the depth fusion, enabling our method to operate robustly under dramatically different, temporally and even spatially varying, measurement conditions.…”