At present, one of the key tasks in ensuring security at sites with large crowds is the identification of potential offenders to counteract terrorism related threats. This is due to the necessity of accounting for a wide range of biometric characteristics that may define a potential offender. In order to develop effective measures aimed at preventing terrorism-related crimes, particularly in the form of armed attacks, it is essential to form an understanding of the offender’s identity – something that is impossible without studying their individual traits. Among the most widely used and highly significant physiological indicators for criminology, law enforcement, and security activities are body movements (such as gait, gestures, and posture) and facial expressions. The primary reasons for the widespread use of these indicators include the ability to perform recognition through surveillance cameras, the inherent difficulty in concealing of disguising gait features, and the potential to extract additional information from low spatial resolution video footage. This study discusses the challenges associated with the fusion of biometric data and the recognition of potential offenders based on body and facial movements captured through video surveillance. It provides a synthesis and systematization of various scientific perspectives – both domestic and international – on addressing the issues of identifying potential offenders by body and facial movements under conditions of incomplete input data and the multifaceted nature of human identity. An architectural framework for a system that integrates biometric features and identifies offenders based on body and facial movements is proposed, with a detailed description of its modules. This framework can serve as a foundation for the development of advanced security control systems. The authors also propose their own approaches to addressing the identified challenges. It is noted that for accurate recognition of potential offenders in the presence of multiple biometric features, it is advisable to employ a combination of data fusion and recognition methods. A hybrid approach is proposed, based on two artificial neural networks and Dempster-Shafer theory for the classification module within security control systems. This is followed by data processing using the random forest method, which is built using an ensemble of decision trees. This structure supports the interpretation of neural network outputs in scenarios characterized by low data accuracy.