Depth Sensors-Based Action Recognition Using a Modified K-Ary Entropy Classifier

Abstract

Surveillance system is acquiring an ample interest in the field of computer vision. Existing surveillance system usually relies on optical or wearable sensors for indoor and outdoor activities. These sensors give reasonable performance in a simulation environment. However, when used under realistic settings, they could cause a large number of false alarms. Moreover, in a real-world scenario, positioning a depth camera at too great a distance from the subject could compromise image quality and result in the loss of depth information. Furthermore, depth information in RGB images may be lost when converting a 3D image to a 2D image. Therefore, extensive surveillance system research is moving on fused sensors, which has greatly improved action recognition performance. By taking into account the concept of fused sensors, this paper proposed a novel idea of a modified K-Ary entropy classifier algorithm to map the arbitrary size of vectors to a fixed-size subtree pattern for graph classification and to solve complex feature selection and classification problems using RGB-D data. The main aim of this paper is to increase the space between the intra-substructure nodes of a tree through entropy accumulation. Hence, the likelihood of classifying the minority class as belonging to the majority class has been reduced. The working of the proposed model has been described as follows: First, the depth and RGB images from three benchmark datasets have been taken as the input for the model. Then, using 2.5D cloud point modeling and ridge extraction, full-body features, and point-based features have been retrieved. Finally, for the efficacy of the surveillance system, a modified K-Ary entropy accumulation classifier is optimized by the probability-based incremental learning (PBIL) algorithm has been used. In both qualitative and quantitative experimental results, the testing results have shown 95.05%, 95.56%, and 95.08% performance over SYSU-ACTION, PRECIS HAR, and Northwestern-UCLA (N-UCLA) datasets. The proposed system could apply to various real-world emerging applications like human target tracking, security-critical human event detection, perimeter security, internet security, public safety etc.

Keywords

• 2.5D cloud point
• full body features
• point-based features
• probability-based incremental learning
• RGB-D
• K-Ary entropy accumulation