A new model for classification of human movements on videos using convolutional neural networks: MA-Net

ABSTRACT

Classification of human movements is essential for interpreting and describing human activities, such as environment-supported living in smart home environments, elderly nursing homes, visual tracking, object tracking, anomaly detection, medical visualisation and mimic analysis. Also, human movements recognition from videos has become one of the important issues that arise with the developing technology and the processing of big data in computers. In this paper, it is aimed to classify human movements by using a data set including different motion videos. For this aim, a new model MA-Net named by us is proposed. MA-Net have 43 layers. In order to examine MA-Net, data having150 videos and 10 classes in UCF dataset is examined. At first, the frames from videos are extracted. In study, it has been worked to take one frame in 50 frames in videos. After that, dataset is classified using well known models such as Resnet50, Alexnet, Inceptionv3, Densenet201 architectures. After, proposed new model MA-Net are classified too. The highest accuracy rate is obtained from MA-Net model as 91.34%.

KEYWORDS:

• Human activity recognition
• deep learning
• CNN
• classification

Highlights

• In this study, video images containing human movements were classified.

• There are 10 different movement classes.

• The results were obtained with both the existing Cnn models and a new model was developed. (MA_NET)

• The new model developed consists of 43 layers. and the highest accuracy value was obtained in the developed model (MA_NET).

Acknowledgments

Authors thank owners of the database for sharing their data;

UCF Sports Action Data Set, https://www.crcv.ucf.edu/data/UCF_Sports_Action.php.

Disclosure of potential conflicts of interest

No potential conflict of interest was reported by the author(s).

Additional information

Notes on contributors

Muhammed Yildirim

Muhammed Yildirim is currently a Ph.D student at Department of Computer Engineering, Firat University. His main areas of work are Deep Learning, Image Processing and Big Data.

Ahmet Çinar

Ahmet Çinar is currently Asst. Prof. in the Department of Computer Engineering at Firat University. His main fields of study are Deep Learning, Image Processing, Computer Graphics, Geometric Modeling and Mesh Generation.