Anthropometric Based Real Height Estimation Using Multi Layer Peceptron Ann Architecture In Surveillance Areas

The work is aimed at automatically learning the real height of the person from single uncalibrated image using anthropometric metric pair cumulative ratios.In this context,application of neural networks to estimate the person real height is explored.A comprehensive data set of 264 images comprising 22 males and 11 females with each of 10 instances are used for experimental study.For feature extraction, we have taken 10 anthropometrics namely stature,neck height,acrominal height,head length,mouth to top of head distance,forehead to chin distance,sellion to chin distance,biocular distance and bitragion distance from full body uncalibrated photograph of a person and computed their pairwise ratios.These ratios are used as inputs to train Multi layer perceptron neural networks(MLP).The performance of model is evaluated using Regression correlation coefficient(R)and RMSE.Comparitive indices of the optimized ANN with input values refering the pair wise rations of Antropometrics for prediction of person real height are R=0.964644 and RMSE=2.484 (males),R=0.8981 RMSE=8.1375(females) and R=.812(males and females together).The results have proven that MLP network is good method to model the accurate real height estimator and could be deployed in distributed surveillance environments for person re-identification in occluded regions with in overlapping and non overlapping camera Field of View(FoV).