Nowadays, the visual information captured by CCTV surveillance and body worn cameras is continuously increasing. Such visual information is often used for security purposes, such as the recognition of suspicious activities, including potential crime- and terrorism-related activities, and violent behaviours. To this end, specific tools have been developed in order to provide law enforcement with better investigation capabilities and to improve their crime and terrorism detection and prevention strategies. This work proposes a novel framework for recognising abnormal activities where the continuous recognition of such activities in visual streams is carried out using state-of-the-art deep learning techniques. Specifically, the proposed method is based on an adaptable (near) real-time image processing strategy followed by the widely-used 3D convolution architecture. The proposed framework is evaluated using the publicly available diverse dataset VIRAT for activity detection and recognition in outdoor environments. Taking into account the non-batch image processing and the advantage of 3D convolution approaches, the proposed method achieves satisfactory results on the recognition of human-centred activities and vehicle actions in (near) real-time.