learning anomalous human actions using frames of interest and decoderless deep embedded clustering

Muhammad Hafeez Javed, Zeng Yu, Tianrui Li, Noreen Anwar, Taha M. Rajeh

Abstract: Inconsistent data and unclear labels make it difficult to learn anomalous behavior from video. Therefore, methods based on deep clustering are now trending in this area. A deep clustering strategy usually relies on encoding and reconstruction to facilitate information discovery. However, it seems pointless to reconstruct the input after the model’s learning process is already concluded. On the other hand, multiple input types carry various features which may help identify the problem more accurately. Hence to mitigate the requirement of utilizing assorted features with clustering, we propose Skeletal Based Autoencoder (SKELBA), which allows us to process the different types of inputs parallelly. The model consists of a spatial graph convolution operator, which helps us convolve the skeletal data more precisely. A decoder-less deep clustering architecture is introduced to enhance the stability of clustering. The relation between reconstruction error and minimizing the lower bound of mutual information (MI) helps us look into decoder-free systems. The joint venture of local–global feature collection and decoder-free encoders techniques shows improved results. Extensive experiments performed on the various benchmark datasets highlight the proposed model’s superiority among recently proposed approaches in the same field.