Human activities and status recognition using depth data

Jiajun Fu, Purdue University


Human activities and status recognition is taking a more and more significant role in the healthcare area. Recognition systems can be based on many methods, such as video, motion sensor, accelerometer, etc. Depth data of Kinect sensor is a novel data type with 25 joints of information, which is popular in motion sensing games. The 25 joints include the head, neck, shoulders, elbows, wrists, hands, spine, hips, knees, ankles, and feet. This paper presents a recognition system based on these depth data. Depending on the characteristic from 5 kinds of statuses, a local coordinate frame was established and partial Euler angles were extracted as features. These Euler Angles can accurately describe the joints distribution in 3 dimensional space. With this kind of feature, a neural network model was built using 50000 sets of data to classify the daily activities into these 5 statuses. Experimental results of 10 volunteers’ action sequences in the same environment showed the accuracy was up to 97.96 percent. The recognition in dark environment had more than 90 percent accuracy.




Chen, Purdue University.

Subject Area

Electrical engineering

Off-Campus Purdue Users:
To access this dissertation, please log in to our
proxy server