A CNN-Based Method Using Optimized Parameters for Dynamic Human Action Recognition

Classification of human action and anomaly detection are important for many areas such as smart cities and security. In the literature, applications are developed on real-time image streams taken from CCTV cameras. Analysis of camera records, classification of human movements and anomaly detection are the motivation of this study. In this study, KTH (Human Action Recognition) dataset is used. Videos in the dataset are divided into frames. There are 6 classes in total, namely Walking, Jogging, Running, Boxing, Hand waving and Hand clapping movements. Keyframes are extracted using 600 videos. A total of 18000 images are used, 3000 frames for each class. These images are collected from 4 different conditions. Labeled images are classified with CNN model. In the proposed CNN model, the parameters of Convolution, Pooling and other layers are optimized and a dynamic model is presented. 96% accuracy is calculated with the proposed CNN model and performance metrics are compared with the literature.