A Multi-batch Differential Binary Motion Image and Deep Hashing Network for Human Action Recognition

Human action recognition (HAR) has garnered considerable interest within the computer science community in recent times, primarily due to its diverse applications in areas such as surveillance and emotion analysis.Classifying human activity takes time and scenes with an ambiguous background. In this paper, a preprocessing stage of the HAR is used to make advantage of the extraction of multiple 2D Differential Binary Motion History (DBMHs) from recordings of human movement spanning spatio-temporal frames. Subsequently, in order to accurately characterize changes in gradient, we generate two distinct types of descriptors for these batches of Directional Binary Multiscale Histograms (DBMHs). This process involves employing both a Histogram of Oriented Gradients (HOG) and a deep hashing network. For building a class-based dictionary, each set of HOG descriptors corresponding to a batch is processed independently. The final view-invariant feature representation is formed by amalgamating the encoded feature vectors from all these batches. Lastly, we refer to the utilization of the hash method, which has recently emerged as the most efficient approach for the retrieval of large-scale images.This study shows how recording changes in deep hash components over time may be used to detect human activity by combining the advantages of deep hashing with deep auto-encoders. The experiments performed on the KTH and Weizman dataset gave our technique its outstanding performance. The results of the study demonstrate that our approach outperforms previously reported methods and reaches accuracy levels of 99.44% and 98.5% with regard to the Weizmann and KTH datasets, respectively.