Full Body Pedestrians Orientation Estimation using Machine Learning

Abstract

In human body orientation estimation (HBOE), body parts play a vital role to build the structure of humans. The estimation task has gotten tremendous attention in previous years due to its applications in almost every field of the real world like medical, surveillance, sports, augmented reality, animations, and many more. Although different approaches for HBOE have been presented, these methods still face obstacles like rapid variation in pose, different viewpoints, camera issues due to weather conditions, etc. The main purpose of this study is to deal with these addressed issues. So much work is being done in this field of research. In this research the dataset called big dataset for body orientation (BDBO) is used. The main purpose of the research is to give high estimation accuracy in short time. The research consists of a few core steps i.e. initial preprocessing, features extraction, features selection and classification. Additionally, the performance of the employed techniques is evaluated and studied to highlight the high results. The first step is image pre-processing in which image sharpening is performed and the resolution of images is enhanced with SRGAN-VGG54. For feature extraction the features from the dataset two CNNs are used in which one is VGG-19, a pre-trained network and the other is a proposed net called BlackNet. The main contribution of this research is BlackNet, which is used to extract the useful features form the dataset for better accuracy. After extracting the features from two CNNs, features are then fused. After feature fusion, features are then passed through the phase of feature selection. For this purpose Whale Optimization Algorithm (WOA) is used for extraction useful and optimal features from fused features. These optimal features are then passed to the state-of-the-art classifiers which are SVM and KNN. A detailed analysis of proposed methodology is given in the sections below to highlight the contribution of this research.