Statistical Features Versus Deep Learning Representation for Suspicious Human Activity Recognition

For many years, researchers have been interested in human activity recognition for many reasons, including preventing crimes before they occur, preventing acts of sabotage, and securing people and facilities. Human activity recognition using deep learning is the focus of many researchers. Techniques for human activity recognition that rely on sensors worn by a person with his limbs and torso are impractical and the only useful way to recognize the human activity especially in public places is through video clips of surveillance cameras. In this paper i compared the performance of CNN with different models such as Inception, ResNet, Inception-ResNet, MobileNet V2, NASNet and PNASNet and the performance of handcrafted features such as statistical features (shape moments like mean, Skew, Kurtosis, etc.) in terms of their human activity recognition accuracy and also compared the results with the state-of-the-art methods. In my research i used suspicious activities included in the HMDB data set (falling to the floor, punching, kicking and shooting a gun) to evaluate the statistical features and also to evaluate the different modern CNN architectures. The experimental results in the case of using statistical features confirmed the superiority of the Support vector machine (SVM) than other classifiers. My experimental results indicated that the CNN with NASNet architecture achieves the best performance of the six CNN architectures but when comparing the performance with the statistical features method, I found the superiority of statistical features with a support vector machine classifier. This paper contributes to studying the effectiveness of using modern CNN architectures in recognizing suspicious human activity and found that these techniques depend on the quality of images, whether the activity is individual or group. This study enables the researcher to evaluate and compare the different modern CNN architectures in suspicious human activity recognition and compare them also with models that use the hand-crafted features represented in the statistical features on an objective and fair basis The remainder of this paper is organized as follows: Section 1 provides the related work. Section 2 describes an action recognition system using Hand-crafted features. Section 3 presents the experimental results and Section 4 concludes the paper.