The main objective of this paper consists of building a reliable and lightweight human faces and actions recognition classifier. It is able to classify a large number of actions, and be lightweight enough that it can work in real time, i.e. process at least 30 frames per second, using of-the-shelf computer hardware, connected to a normal CCTV infrastructure. TCN represents a viable solution to the proposed problem. It can classify a large number of actions - 60, using only RGB images of fairly low resolution, in real time.Deciding which class an action belongs to should not be tied to a particular environment, person, background, view angle, or other specific identifiers. This decision should be tied only to the person executing it, and the spatial-temporal context in which the person is present. As technology and processing power improve, the problem slightly shifts. When more processing power to a system is added, is possible either to increase the number of frames per second or the number of cameras in the infrastructure, or to increase the quality of the images, resulting most likely higher accuracy of the predictions. The model can also be extended to a larger number of classes, with a minimal impact on performance.The proposed model has an accuracy of 82%, which can be attributed to the recurrent property of the network. The model performs close to the state-of-the-art level.The TCN + 3D Convolution Model is built with the smaller TCN units. Its architecture uses an alternation of a Simple Unit 5a and a Complex Unit 5b, in order to maximize the diversity of features the model learns. After one simple and one complex unit, there is an Average Pool 3D layer that reduces at least one dimension to half. The image starts with 20 frames, 3 channels and resolution of 192x108. Depth is arguably the most important dimension. It provides context to the convolution layers, and determines the width of the temporal frame in which the convolutions are applied. Finally are mentioned some results, conclusions and references.

TCN Units, Solution in Recognition of Human Activities

Paper Proceedings: https://asmedigitalcollection.asme.org/ISPS/ISPS2021/volume/84799

ASME 2021 Information Storage and Processing Systems (ISPS2021)

Supporting the need among researchers worldwide for additional online and multimedia content, ASME now offers the ASME Conference Video Collection. This platform will collect and provide long-term access to conference presentation videos, encompassing all technology areas within ASME's technical conference and events program. About 30,000-40,000 engineers across all mechanical engineering and related disciplines participate in ASME conferences each year. This includes authors of approximately 6,000 to 8,000 technical papers published by ASME, the majority of whom also submit video presentations for our virtual conferences. The video collection will be available as an add-on to existing product packages or as a separate subscription.