Activity recognition
A variety of techniques are used automate the collection and classification of workout data gathered by a wearable physiological monitor. The classification process is staged in order to correctly and efficiently characterize a workout type. Initially, a generalized workout event is detected using motion and heart rate data. Then a location of the monitor on a user is determined. An artificial intelligence engine can then be conditionally applied (if a workout is occurring and a suitable device location is detected) to identify the type of workout. In addition to improved speed and accuracy, a workout detection process implemented in this manner can be realized with a sufficiently small computational footprint for deployment on a wearable physiological monitor.
1. A computer program product comprising computer executable code embodied in a non-transitory computer-readable medium that, when executing on a wearable physiological monitor, performs the steps of:
receiving data from a number of sensors on the wearable physiological monitor, the data including accelerometer data and heart rate data acquired by the wearable physiological monitor;
applying a threshold based on at least one of the accelerometer data and the heart rate data to identify two endpoints of an interval of increased physical activity indicative of a workout by a user of the wearable physiological monitor;
dividing the data including the accelerometer data and the heart rate data into a number of sequential segments;
determining a probability that each one of a number of sequential segments of accelerometer data includes data from one or more locations on a body of the user by applying a machine learning algorithm to each one of the number of sequential segments;
selecting one of the one or more locations having a highest overall probability of being a current position from all of the number of sequential segments as a position of the wearable physiological monitor on the body of the user; and
conditionally employing an automatic workout classification algorithm to detect a type of the workout only when the position is an ankle or a bicep of the user, wherein the automatic workout classification algorithm includes a deep convolutional neural network trained to calculate a probability that a chunk of data including at least one of accelerometer data and heart rate data from the wearable physiological monitor during the workout is each of a number of candidate types for the workout.
2. The computer program product of claim 1 further comprising code that performs the step of conditionally employing the automatic workout classification algorithm only when the position is the ankle of the user.
3. The computer program product of claim 1 further comprising code that performs the step of conditionally employing the automatic workout classification algorithm only when the position is the bicep of the user.
4. The computer program product of claim 1 wherein the automatic workout classification algorithm detects the type of the workout based on a history of exercise for the user.
5. The computer program product of claim 1 further comprising code that performs the step of applying the deep convolutional neural network to a number of chunks of data from the wearable physiological monitor to obtain a posterior distribution of the number of candidate types for the workout.
6. The computer program product of claim 5 further comprising code that performs the step of determining the type of the workout by selecting one of the number of candidate types in the posterior distribution having a highest probability of characterizing a workout type for the workout.
7. The computer program product of claim 1 further comprising code that performs the step of updating the automatic workout classification algorithm on a central server based on new data from a plurality of users.
8. The computer program product of claim 1 further comprising code that performs the step of adapting the automatic workout classification algorithm to a specific user based on prior workout data for the specific user.
9. The computer program product of claim 1 further comprising code that performs the step of determining whether a different position for the wearable physiological monitor can provide more accurate data for the type of the workout and, when the different position can provide more accurate data, providing a notification to the user suggesting a movement of the wearable physiological monitor to the different position.
10. A system comprising:
a wearable housing;
one or more sensors in the wearable housing configured to provide heart rate data and accelerometer data for a user of the wearable housing;
a memory storing a machine learning algorithm that estimates a probability of a location of the wearable housing on a body of the user using data from an accelerometer; and
a processor configured by computer executable code stored in the memory to identify a workout by the user based on at least one of the heart rate data and the accelerometer data, determining a probability of a position on the body of the user for each of a number of sequential segments of the accelerometer data by applying the machine learning algorithm to each of the number of sequential segments of the accelerometer data from the workout and selecting the position on the body of the user based on one of a number of candidate positions having a highest overall probability of being an actual position of the wearable housing for the number of sequential segments, and conditionally employ an automatic workout classification algorithm to detect a type of the workout when the position is an ankle of the user or a bicep of the user, wherein the automatic workout classification algorithm includes a deep convolutional neural network trained to calculate a probability that a chunk of data including at least one of accelerometer data and heart rate data from the wearable physiological monitor during the workout is each of a number of candidate types for the workout.
11. The system of claim 10 wherein the processor is configured to conditionally employ the automatic workout classification algorithm only when the position is the ankle of the user.
12. The system of claim 11 wherein the processor is configured to create a notification to the user to move the wearable housing to the ankle of the user when the position is not the ankle of the user.
13. The system of claim 10 wherein the processor is configured to conditionally employ the automatic workout classification algorithm only when the position is the bicep of the user.
14. The system of claim 13 wherein the processor is configured to create a notification to the user to move the wearable housing to the bicep of the user when the position is not the bicep of the user.
15. The system of claim 10 wherein the processor is configured to apply the deep convolutional neural network to a number of chunks of data from the one or more sensors to obtain a posterior distribution of the number of candidate types for the workout.
16. The system of claim 15 wherein the processor is further configured to determine the type of the workout by selecting one of the number of candidate types in the posterior distribution having a highest probability of characterizing a workout type for the workout.
17. The system of claim 10 further comprising a central server, wherein the processor is configured to update the automatic workout classification algorithm on a central server based on new data from a plurality of users.
18. The system of claim 10 further comprising a central server configured to adapt the automatic workout classification algorithm to a specific user based on prior workout data for the specific user.
19. The system of claim 10 wherein the processor is configured to determine whether a different position for the wearable housing can provide more accurate data for the type of the workout.
20. The system of claim 19 wherein the processor is configured to, when the different position can provide more accurate data, provide a notification to the user suggesting a movement of the wearable housing to the different position.