Its not the final posture we wanted to check but rather the process that is leading you to that posture. For this we needed to make the process more dynamic which allowed us to compare frame by frame of the video showing the correct posture and the video of the exercise we are doing. The reason we did this is because even if we are able to achieve the final correct posture the process may not be right and with the previous model we didn't have a way of detecting that. Thus we worked upon frame detections based on that. Doing that we compared the score of the individual frames of both videos which gives us an indication whether or not the user is following the process correctly or not.
We were able to successfully replicate the process of comparing videos frame by frame and giving the final output as a series of scores for different video frames. The average of all the scores was also taken in order to depict the user with what average accuracy is the user doing the exercise. A major problem with this work is the preprocessing if the data. The video needs to be short and clipped so as to only focus during the exercise part otherwise the output data would be too noisy to make decisions from. We will be working on this problem statement in the future so as to give an effective metric system to the user.

We tried to create our own dataset because this type of dataset was not available to us. This took us a long time to create because of the space and time restrictions.