Yoga is an ancient practise that has been followed to build up our physical and mental fitness. Yoga has been a form of treatment in clinical therapy, improving immunity and assisting the recovery of various chronic conditions. If performed incorrectly in the absence of a yoga expert, there is a risk of injuries such as strains, sprains, slip discs. Hence, there is a need for accurate real-time yoga posture classification techniques to be implemented that can be used as part of any automated yoga posture monitoring applications. This paper aims to perform Deep Learning based automated recognition of yoga asanas in real time from two different views accurately. We prepared a yoga posture custom video dataset titled "DIAT-YogaDS" consisting of five unique asanas from “Suryanamaskar”, a modern-day commonly followed yogic practise. These unique asanas contain occluded portions, and it is found that occlusion sensitive Human Pose Estimation (HPE) models (OpenPose, Detectron2, and YOLOv8-Pose) fail to recognize all key points, while DensePose HPE takes more time compared to MoveNet HPE to extract these features. In this work, to achieve real-time performance, we used MoveNet to obtain joint vector cosine angle features from extracted skeletal body key points. The proposed Deep Neural Network (DNN) model trained on these angle features is lightweight, providing robust detection invariant to both scale and distance. Experimental results show the precision of the proposed DNN to be 0.9428. The proposed DNN model can be implemented on edge devices, and it requires very low computational resources.