| Active triangulation is a well-established technique for collecting range points. Low-power systems with limited computation resources are desirable because of price considerations, portability and eye safety requirements. Such systems present challenges to correct operation, especially in environments with high ambient light, or when requirements specify a high ratio of maximum to minimum operating range. This work makes contributions in three areas to help meet these challenges.;First exposure control is examined. A photometric analysis of relative irradiance expected at the camera sensor is performed. The limiting effects of eye safety compliance, minimum realizable shutter times and pixel bit depth for linear response cameras are considered. Quantitative results are determined for dynamic range requirements on the camera, predicting when laser return can be expected to produce the brightest image pixels, and when exposure control is needed. We introduce an algorithm to estimate optimal exposure parameters from analysis of a single, possibly under- or over-exposed, image. A new quantitative measure of exposure quality is presented for use with this algorithm, based on average rendering error due to quantization. To compute the exposure quality in the presence of saturated pixels, we propose fitting a log-normal brightness distribution using a right-censored estimator. Experimental results are presented comparing the estimated versus "ground truth" optimal exposure parameters.;Second, we consider the problem of return detection under high ambient light conditions. A detection method combining filtering of specialized low-level features with dynamic programming is proposed. Optimal parameters are selected from a precision-versus-recall Pareto front formed from a grid search of the parameter space. Results are evaluated against ground truth and can approach results achieved using current techniques under ideal conditions.;Third, real-time model fitting with limited computation resources is considered. A 'summation-difference' implementation of the segmented least squares algorithm which runs in O (n2) time, where n is the number of input points, is applied. Model fits generated using variants of this algorithm and ransac are evaluated on test images using two novel methods, one of which is an adaptation of Needleman-Wunsch sequence alignment. This evaluation enables optimal tuning of model fitting parameters. |
