Event recognition [1] is a technique to identify in a video stream a semantic action, for example “start walking”. The pixels on the screen are doing something, and the activity parser recognizes the action. He converts the low level binary information into a high level natural language description. The detected events are stored in a log file which makes it easy to search for certain words, for example to query for all video-cllips in which a person is walking.

The problem is, that for robotics applications such an event detection system is useless. Because recognition alone is different from replay an action. The problem of a robot isn't to observe a human who is doing a task, but the robot should act by it's own. The underlying model which is able to detect events, actions and long term behaviors isn't completely useless, because it goes into right direction. It helps to reduce the state space. The problem is to build on top of an event parser a robot control system.

Suppose, an event detection is available which can recognize basic actions like stand, walk and run. How can such a model be transferred into a motion controller which will produce the servo commands for a robot?

[1] Mo, Shiwei, and Daniel HK Chow. "Accuracy of three methods in gait event detection during overground running." Gait & posture 59 (2018): 93-98.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.