Compound human action recognition from low illumination video under different ambient environment using hybrid cnn algorithms

Abstract

Compound Human Activity Recognition (CHAR) plays a vital role in various applications such as surveillance, terrorists activity monitoring. CHAR is detected from video images. CHAR has many challenging problem in recognition because of noise in video images such as human body shape and motion variation, occlusion, cluttered background moving cameras elimination conditions view point variations, low light illuminations videos. Low light illuminations videos have the objects in the image with low or high contrast, brightness, saturations, shadow warmth and sharp. However, occlusions in image, moving camera viewpoint variation are solved through various traditional algorithms such as KNN, SVM and CNN. The major problems addressed in this thesis are (i) a robust system to recognize the compound human action with less classification errors and improve the efficiency of recognition. (ii) Improved classification performance need to be achieved with reducing the classification errors. Moreover, CHAR from low illumination video with multiple problems in a single frame such as low contrast, saturation, brightness, warmth, and sharpness need an efficient algorithm. In this thesis, Compound Human Action recognition has classified as HHOR.OFHR and HAR. In HHOR the various actions such as (i)Eating (ii) Phoning (iii)Drinking are considered. In OFHR, the action such as (i)Hand Waving (ii) Handshaking (iii) Pulling (iv) Pushing are considered for recognition. In HAR the following actions such as (i)Walking (ii)Running (iii) jogging are considered. newline