FitNet: A deep neural network driven architecture for real time posture rectification

A methodology of real time pose estimation, which is believed to mitigate many orthopaedic adversaries pertaining to wrong posture, has been illustrated in this paper. Vast array of problems get reported that are known to arise due to maintaining a wrong posture during exercising or performing yoga, for a prolonged period of time. Several developments were made with regard to this issue, yet a major drawback was the presumption that a person during exercising or performing yoga or any kind of gym sessions, will keep the camera facing only at a fixed pre-determined portrayal direction. The approach, towards this problem, mainly deals with precise ROI detection, correct identification of human body joints and tracking down the motion of the body, all in real time. A major step towards converging to the solution is determining the angular separation between the joints and comparing them with the ones desired. Another important facet of the stated methodology is analysis of performance of the deep neural architecture in different camera positions. This is a major bottleneck for many different models that are intended to track posture of a person in real time. All these operations are done efficiently, with an appropriate trade-off between time complexity and performance metrics. At the end a robust feedback based support system has been obtained, that performs significantly better than the state of the art algorithm due to the precise transformation of input color space, contributing significantly in the field of orthopaedics by providing a feasible solution to avoid body strain and unnecessary pressure on joints during exercise.