With old age & senility, it gets harder to count reps.  There is of course the ages old

https://www.amazon.com/GOGO-Counter-Carnival-Manual-Mechanical/dp/B001KX1VW2

counter from grumpy old school bus driver lore.  Maybe it was too rooted in grumpy bus driver lore.  It has really small digits which are hard to see in old age & it might not have enough tactile feedback.

The next idea was to convert a board from the day job into a counter.  They have big displays.  It would need a haptic engine, also known as a vibrator in the days when the world was manely married.  The problem is it's not convenient to look at a paw while exercising.

The next idea was a wireless button with haptic engine that incremented a counter on a phone or computer screen.  This would take a while to set up, every day.  

Maybe another remote control with a haptic engine for Heroineclock II could be devised, to turn it into a giant counter.  It might also be good enough with some kind of sound effect.  You still have to carry something around & point it at the clock & you can't tell time while counting.

A giant counter still sounds appealing, for some reason besides counting reps.  20 years ago, a national debt clock took only 13 digits.  Today, it takes 15 digits.  It could be used for cooking.

To be completely free from looking at a paw, complex setup, carrying something around, & generating haptic feedback, you really need a machine vision program that detects poses.  There would be 1 pose to reset the counter.  Then, it would detect 2 poses from each exercise.  The number could still be displayed on another screen.

It's surprising that for all the programs which now use only speech as user input, there is no significant marketing of pose detection.  There were a few quad copters which used poses, but they weren't widely marketed.

Besides machine vision, there is strapping wireless accelerometers or the infamous mocap balls to the 4 legs.  This has an involved setup & wouldn't be easy to reset the counter with.

There is source code for pose detection:

https://github.com/llSourcell/3D_Pose_Estimation

It's based on a database of 3.6 million human poses that were captured with both mocap & video.  The neural network simply compares images to get the closest pose & looks up the corresponding mocap data.  It's very slow.  The mane emphasis is analyzing crowds.

There's a pose estimator that works in a browser.

https://medium.com/tensorflow/real-time-human-pose-estimation-in-the-browser-with-tensorflow-js-7dd0bc881cd5

An openCV implementation:

https://www.learnopencv.com/deep-learning-based-human-pose-estimation-using-opencv-cpp-python/#comments

A summary of pose estimators:

https://becominghuman.ai/human-pose-matching-on-mobile-a-fun-application-using-human-pose-estimation-part-1-intro-93c5cbe3a096

It's come a long way since lions 1st studied it, 6 years ago, yet there is no pose estimator app for a phone.

The holy grail seems to be openpose.  This is what all the tracking cameras & quad copters in the last year are using for tracking subjects.

https://github.com/CMU-Perceptual-Computing-Lab/openpose

Counting exercise poses is much simpler than detecting the absolute position of each body part from a photo.  The neural network could probably be trained with a few reps of each exercise.  A stereo camera system might even be easier.    The kinekt did this before it was discontinued.