A Deep Convolutional Neural Network (DCNN) and Squirrel Search Algorithm (SSA) based Classifier Framework by Extracting Human Body Skeleton points based on Silhouette Images for Human Action Recognition in Image Processing

Abstract

Due to the development of cost-effective depth sensors and rapid poses estimation algorithms, skeleton-based action recognition is widespread. Historical approaches based on pose descriptors often fail with large-scale data sets as engineered features are limited. In this paper we intend to reinforce the geometric connections between joints to identify behavior. Three basic geometries are incorporated: joints, edges and surfaces. For action detection, the DCNN-based network uses a novel perspective transition layer and time dropout layers to learn robust images. Consequently, we propose using the Squirrel Search Algorithm (SSA) algorithm in order to make the Deep Convolutional Neural Network (DCNN) classification more efficient. In this report, we extract human body skeleton based on silhouette images by using a distance gradient, to classify crucial points from silhouette images that play a significant role in the recognition of human activity. Experiments of 3d large-scale measuring identification datasets show that joints, edges and surfaces for specific behavior are efficient and complementary. Our solutions greatly exceed present state-of-the-art strategies for identifying functions.