The human actions recognition in video is a topic of growing interest in the scientific community in computing, due to its application in real problems and different domains such as video surveillance, medicine, psychiatry, among others, and on the other hand, due to the overcrowding of video capture devices all over the planet. Processing video to extract characteristics and subsequent classification or recognition is a complex task, as it involves processing data in a spatial dimension (video dimensions) and a temporal dimension, causing the input data to increase abundantly and become in a challenging task. There are two approaches to the recognition of human actions on video; handcrafted approaches based on optical flow and approaches based on Deep Learning, the latter has achieved many achievements in terms of accuracy; however, it has the problem of high computational cost, making its application almost impossible in specific domains, much less in a real-time scenario. In this way, we propose an architecture based on Deep Learning, for human actions recognition in video, oriented to the domain of video surveillance and in a real-time scenario; For this, the proposal is based on an architecture that combines 3D CNN and DenseNet techniques. The results show that the proposal is efficient and can be used in the domain of real-time video surveillance. Likewise, general representations are proposed referring to the resolution and minimum frames per second that guarantee recognition.