Towards generalised time-of-flight range imaging at the edge of moving objects

Classifying Transverse Motion in Time-of-Flight Range Imaging

“…Random forest performed better in the range 0.75-0.80, but at the cost of a very large tree. The interesting feature is that the proposed method appears robust to noise, an advantage over previous work [2] which rapidly declined in performance with noise level. Testing on the independent set saw a small a drop in performance, indicating the possibility of over fitting in cross validation.…”

“…Explicit modelling and system inversion can determine the range, provided that the step change at the pixel is both detected and localized to the correct phase step. Analysis of the second Fourier bin reveals the presence of a step change [2] but does not localize it. In this work we show how to localize the step change to phase step n.…”

“…When the edge of an object in transverse motion passes over the pixel field of view, α, β , and φ can change [2]. Explicit modelling and system inversion can determine the range, provided that the step change at the pixel is both detected and localized to the correct phase step.…”

“…θ n = 2π(n − 1)/N with N = 9. For step change at each n, 1001 simulations were taken, with random φ , α, and β and with random step values of φ and α (β is assumed constant [2]). The simulation was repeated for additive zero mean Gaussian noise with standard deviation of 0%, 0.5%, 5% and 10% of the average value of α.…”

“…Here we are most concerned with the motion of an edge over a pixel, which causes a step change. Explicit modelling of the step change can lead to error free imaging [2], provided that the point in time that the step change occurred is correctly identified.…”