Person re-identification is at ask that aims to identify a subject from a video sequence captured by multiple cameras with a non-overlapping field o f view. It has many applications in surveillance, public safety, smart cities and especially in context-aware holographic communication. The multiple frames in a video sequence can be used to find temporal r elations and t heir representation can be utilized for improving the re-identification task. Furthermore, semantic information related to the person can be used to improve the performance and deliver better representations of the video sequences and images. In this work, we propose a novel approach for the utilization of semantic information in learning non-local weights by capturing the relations among features extracted from video sequences. The semantic information is represented by the human body and extracted using a dedicated neural network. Experimental validation using popular datasets confirms the validity of the proposed approach.