Comparative Analysis of Pre-trained based CNN-RNN Deep Learning Models on Anomaly-5 Dataset for Action Recognition

Action recognition in videos is one of the essential, challenging and active area of research in the field of computer vision that adopted in various applications including automated surveillance systems, security systems and human computer interaction. In this paper, we present an in-depth comparative analysis of five CNN-RNN models based on pre-trained networks such as InceptionV3, VGG16, MobileNetV2, ResNet152V2 and InceptionResNetV2 with recurrent LSTM units for action recognition on Anomaly-5 dataset. The performance of these models is analyzed and compared in terms of accuracy, precision, recall & F1-scores and computational efficiency. The CNN-RNN architectures we considered for analysis in this paper, the ResNet152V2 based CNN-RNN model exhibits better performance and achieved highest accuracy, precision, recall and F1-score equal to 92.20% due to its ability to capture more complex spatial features. This comparative analysis may guide the researchers in selecting appropriate models for real-world applications for action recognition. In addition of this, a new dataset is developed called Anomaly-5 that can helps as a valuable resource for training and evaluating action recognition algorithms.