I'm currently prototyping an algorithm for activity recognition from wearable devices such as IMU sensors. I would like to know the desirable accuracy in real-time applications, provided the computational cost is very low. I have read in a paper it around 90% is acceptable in general case. But i don't think its possible for time critical applications, especially wearable devices with little processing power(almost zero lag). I would like to consider mission critical applications as well. I would like to be redirected to any paper related to that, what's the industries requirement. I came close to 85%. Is that enough?
It depends on the application that you are dealing with. Some applications require higher accurac as possibily wrong predictions will cost a company a lot.
Hi Arunraj,
If you're trying to recognize a particular event from the pattern of data you collect with your wearable, you'll need to measure both sensitivity and specificity. Sensitivity is correctly recognizing the event/activity when it occurs. Specificity is correctly identifying when it doesn't occur. Acceptable values of these will depend on what you're measuring, how important the event is, and what you intend to trigger as an action. In some cases, for example, you may be able to live with a number of false positives (lower specificity), but need to be sure that you correctly recognize an event when it does occur. I don't think there are standard values as it's very dependent on the area of application and the impact of measurement and measurement errors.
Good luck with your work.
Bill
Always it depends about the problem you are dealing. If it is important to have high precision because there is life related or maybe you just need to classify some things like if someones is walking or not.
It is depends on ur application. If incorrect prediction is critical in the application, you need to make your application much reliable than before.
- Badges
- Science topic