Recognition of human activities from still image using novel classifier

Abstract

The quest for recognizing human activities and categorizing their features from still images using efficient and accurate classifier is never ending. This is more challenging than extracting information from video due to the absence of any prior knowledge resembling frames stream. Human Activities Recognition (HAR) refers to computer identification of specific activities to aid understanding of human behaviors in diversified applications such as surveillance cameras, security systems and automotive industry. We developed a novel model for classifier and used it in three main stages including preprocessing (foreground extraction), segmentation (background subtraction) to extract useful features from object and sort out these features by the classifier (classification). The model is further simulated using MATLAB programming. Our new classifier generates slightly different results for still image based on dataset INRIA and KTH for 780 images of (64*128) pixels format obtained from literature. The recognition rate of 86.2% for five activities such as running, walking, jumping, standing and sitting achieved by us is highly promising compared to the existing one of 85% over last decade.