Hybrid Adaptive Deep Learning Based Human Action Recognition

Abstract

Many Human Action Recognition (HAR) systems are used to recognize everyday human behaviors such as standing, walking, sitting, running, cooking, and exercising. The existing human action recognition technology is not perfect. It has poor adaptability and low accuracy. Deep learning is used to improve action recognition accuracy and motion detection. Deep neural network models can be difficult to train, such as gradient disappearance, gradient explosion, and over-fitting. This paper will help to reduce the difficulty in training the above deep neural model parameter initialization method based upon the activation function for the multi-layer networks. It also addresses the problem of model training deep neural networks. The presented hybrid algorithm Improved Spotted Hyena algorithm and Seagull Optimization Algorithm, (ISHO-SOA), is used to solve optimization algorithms and obtain minimum optimum values. The segmented image is further classified using different inputs by Adaptive Deep Convolution Networks (ADCNN), with feature maps. The classification method provides performance metrics such as Accuracy and Precision, F1 measure, along with Recall that can be used to recognize the most accurate results and the best classifiers. This algorithm is highly efficient in improving classification accuracy and solving global optimization problems. Experimental results show that the proposed algorithm has a 98.21% higher accuracy than other meta-heuristic algorithms.