i want to recognize human activities in multiple camera environment. I am taking 2 camera views for experiment. I want to fuse information extracted both views to get a precise feature vector. But i am facing following confusions:
As i am using supervised learning, I have to label activities of each person in each frame. In first camera view at some time t, two persons are observed as very close to each other so i label it as interacting. But in second camera view at same time t,it seems that those persons are not close to each other i.e distance is high.. so i labeled it as non- interacting.
How can i fuse two feature vectors (from 2 views) having 2 different labels at same time?
I suspect that associative|statistical machine learning techniques alone will be insufficient for this task. Semantic|conceptual techniques that provide situational context may be of assistance. Automated human activity recognition is a very hard problem. You might also consider alternative feature vector components. For example, activity duration and volumetric features over time. Regardless, some degree of temporal modeling will be required. Good luck!
Yes i know that contextual information is required to accomplish this task...but i want to ask about fusion process
Hi Eman,
Several approaches can be considered in your case, such as taking the mean of those features, to be the resulting fused feature.
HTH.
Samer
Then how i would label the resultant features for svm training? By looking at camera1 or camera2?
Resulting feature should be labeled according to the real situation. Meaning that, you need to do labeling based on the real time observation; if you believe that camera 1 has captured the real situation then label the fused feature based on it, otherwise label it based on camera 2.
HTH.
Samer
- Badges
- Science topic