The Scientific World Journal

Research Article

Domain Adaptation for Pedestrian Detection Based on Prediction Consistency

The proposed domain adaptation algorithm.

Input:
Source domains , where , target domain ,
The maximum number of iterations and a base learning algorithm .
(1) Total training data: .
(2) Initialize the weight , for .
(3) for to do
(3.1) Train a weak hypothesis on with weights .
(3.2) Compute the weighted error of on .
, where is an indicator function.
(3.3) Choose .
(3.4) for to do
(3.4.1) Train a strong classifier on with weights based on standard Adaboost algorithm.
(3.5) Choose for based on decision consistency.
(3.6) Update the weights for all :
(*)
where normalizes to be a distribution.
(3.7) end for
Output: the hypothesis .