Violent behaviour, especially using handheld weapons, is a significant issue plaguing society in this era. This paper evaluates the state-of-the-art 3DCNNsl violent activity recognition architecture for violent activity recognition with an overview of its operations via transfer learning, adapting the 3DCNN pre-trained model for new undertakings. 3DCNNsl violent activity recognition final layer modifications evaluate the action’s generic status, emphasising the activity’s true nature, belonging to the violent or non-violent class with subclasses. The idea emphasises 3DCNNsl violent activity recognition capability and confidence in its predictions regarding the homogeneousness of violence and hostile yet non-violent conduct from an overall perspective. The operation adds value to the violent activity recognition domain by evaluating action similarity data, which intensifies the computational load identified by overall accuracy performance. To establish the model’s performance impact, we employed real-world conditions to observe processing effectiveness on specific violent/non-violent classes concerning their complexity. We designed the conditions to reflect pre-processed data (data containing resolution/scenery enhancements), no pre-processing (raw data without modifications or enhancements), and action similarity (intense complexity between violent/non-violent actions) to evaluate the lethal sporadic complex nature of violence honestly. The designed conditions are only applied to evaluate 3CNNsl violent activity recognition models. Experimental results project the model’s overall accuracy capability at 72–88% utilising data with pre-processing (data containing resolution/scenery enhancements). We concluded the investigations by discussing operational challenges and measures to enhance the model’s effectiveness to produce more robust outcomes.