The detector is inspired by the lightweight BlazeFace model, used in MediaPipe Face Detection, as a proxy for a person detector. It explicitly predicts two additional virtual key points that firmly describe the human body centre, rotation and scale as a circle. The landmark model in MediaPipe Pose predicts the location of 33 pose landmarks.
Since the Mediapipe API gives 33 landmarks, we will use these landmarks to get the details about the position of each body part. We are looking for landmarks of the hip, knee and ankle as per the requirement. Now since we have two legs and this exercise will mostly be done sideways i.e. the camera will be facing the side of the person so we have to choose which leg is in front and which is in the back. Further down the line, I have created logic which tells the program to increase the counter if the user has completed their rep.
On the given testing data the program seems to be working fine. Although sometimes it starts the timer off time due to noisy data points provided by the Mediapipe API. For this, we can use exponential smoothing. Furthermore, we can remove the effect of redundant frames with the help of exponential smoothing. If the change is greater than the threshold the program will retain the previous value but if the change is small the program will first smooth it to then make sure that the angle is increasing gradually because the leg, in this case, will be moving slowly.
-
RUN the following command in Command-Line
pip install -r requirements.txt -
python app.py -
Now you can visit [http://localhost:5000]{.ul} to access the result.
Note: It may take a few seconds for mediapipe to start itself as it needs to start its underlying tflite API.