US 11,927,601 B2
Persistent two-stage activity recognition
Diyan Teng, Santa Clara, CA (US); Junsheng Han, Los Altos Hills, CA (US); Raehan Ahmed Syed, Chicago, IL (US); and Rashmi Kulkarni, Redwood City, CA (US)
Assigned to QUALCOMM Incorporated, San Diego, CA (US)
Filed by QUALCOMM Incorporated, San Diego, CA (US)
Filed on Aug. 31, 2021, as Appl. No. 17/463,451.
Prior Publication US 2023/0065124 A1, Mar. 2, 2023
Int. Cl. G01P 13/00 (2006.01)
CPC G01P 13/00 (2013.01) 13 Claims
OG exemplary drawing
 
1. A processor-implemented method, performed by at least one processor, the processor-implemented method comprising:
receiving, by a first prediction model, sequential data from a sensor related to an activity of one or more of a human or an object;
determining, via the first prediction model, a change in an activity state based on an estimate of a log-likelihood ratio (LLR) function for the sequential data; and
transmitting, to a second prediction model, an indication that the activity state has changed, the indication triggering the second prediction model to wake from an inactive state and determine an updated activity state of the one or more of the human or the object based on the sequential data;
wherein the change in the activity state of the one or more of the human or the object occurs when the estimate of the log-likelihood ratio function is greater than a threshold parameter, which is adjustable to achieve a tradeoff between a false alarm rate and a detection latency.