Hi, I am from Experts Vision and we have been doing research projects similar to this one for over 4 years, like complex activities detection in videos and facial expression recognition etc, you can see our uploaded portfolio for complete list. We have only just entered the freelancing.
For your project, this is what we propose: First , we will apply background subtraction to all action images in the training set. Then we will extract the silhouette.
Next we will use Principal Component Analysis to extract 28 feature vectors from each silhouette.
These feature vectors will be used to get a covariance matrix. This covariance matrix will then be the input to our classification model.
The covariance matrix is simple to compute and has low storage requirements, so it will ensure less execution time resulting in efficient code.
For classification, we will train a Neural Network model (or SVR or K- nearest Neighbor model, if you prefer) and then train it with each training image.
Finally, we will test it for unseen image data. The preprocessing steps for testing data will be the same as for training data.
As for the database, if you want we can construct a database of our own as we have successfully done in the past.
Regards,
Hafsa Asad.
Experts Vision.