Research On Shoeprint Image Based Human's Height Prediction Methods

Footprint is the marks left on the ground or other carriers by people as they walk.Footprint can be divided into three categories including barefoot footprint,footprint of wearing socks and footprint of wearing shoes(shoeprint).Studies have shown that the skeleton of adults does not change much under normal circumstances and is relatively stable in dynamic morphological characteristics when walking,so the footprint is unique and stable.Therefore,using footprint can identify person's ID and estimate their height,gender,age attribute,which are mainly used in the fields of health,security and criminal investigation.In the field of criminal investigation,shoeprint is one of the most important evidence at the scene of a crime.Estimating the height of a suspect based on the shoeprint can narrow the range of suspects and even identify the suspects.Therefore,the research of human height prediction based on shoeprint image has great research significance and application value.At present,the research on height prediction based on footprint image mainly focuses on the barefoot footprint image,and has achieved a good accuracy.However,most of the footprint left at the crime scene is shoeprint.Therefore,this thesis mainly studies the prediction method of human height based on shoeprint image,and the main contents include:(1)Construct a shoeprint image dataset.In this thesis,the shoeprint of different person with different heights is collected to build the dataset.Specifically,the height and shoeprint image are collected from 1157 adult males and 256 adult females,in total of 5652 shoeprint images.For each person,it is collected about 4 shoeprint images on average.Meanwhile,some necessary preprocessing procedures including angle calibration,centralization and processing for fragmentary shoeprint image are carried out for the constructed dataset.(2)Propose a height prediction method combining geometric features and multivariate Gaussian classification.The method takes length,width of shoeprint image and distance between forefoot and backfoot centroid as geometric features These geometric features of shoeprint image are extracted.The person's height is divided into different categories by taking 1cm as interval.The probability of the predicted height belonging to each category is calculated by multivariate Gaussian classification according to the geometric features of the shoeprint images,and the final predicted height is obtained by weighting each category height with respect to its probability.By using proposed method,it achieves 3.84 cm MAE metric.(3)Propose a height prediction method by combining multi-scale local features and global features.The method develop a two-stream network which is consist of a Feature Aggregation Network(FAN)and a Global Structure Network(GSN),these two different networks can capture local and global feature for height prediction.Firstly,the dataset is enlarged by using data augmentation,then the multi-scale local feature and global feature are extracted from FAN and GSN correspondingly.Then,the extracted multi-scale local feature and global structure feature are concatenated to obtain the final feature representation.And finally the final feature representation is fed into a regressor to predict person's height.By using proposed method,it achieves 2.05 MAE metric.The experimental results show that the height prediction method combining geometric features and multivariate Gaussian classification proposed in this thesis is superior to the existing height prediction methods based on geometric features.The height prediction method based on the combination of multi-scale local features and global features is more accurate.However,which is deep learning based method and needs a large amount of data to support.Both methods are suitable for different occasions and have good application prospects.