Human Motion Recognition using OneClass SVM (support vector machines)

I propose a motion recognition method using OneClass SVM (Support Vector Machines). The videoclips in the database are composed of frames on which I applied a filter called "Recursive filter". So the dataset is composed of filtered images. Using principal component analysis (PCA) the feature of human motion is extracted and then the OneClass SVM classifier is employed to classify the motion pattern which is in our case "WALKING" because we are using one class.

The method is based on recursive tracking and reducing noise in order to improve the tracking capability.

matrix_train = None
for image in os.listdir('/home/ahmed/Desktop/dataset1/train'):
    imgraw = cv2.imread(os.path.join('/home/ahmed/Desktop/dataset1/train', image), 0)
    imgvector = imgraw.reshape(160*120)
    try:
        matrix_train = np.vstack((matrix_train, imgvector))
    except:
        matrix_train = imgvector

There are 3 status : "UNOCCUPIED" , "WALKING" and "THREAT" if the new data is different from the training set (which is walking)

This is the classification scheme where we can see the training (yellow dots) and testing (red dots) sets retain the same proportion and the the accuracy level of prediction is about 0.899/1.

As an Improvement, I have to add more classes to recognize more motions and push the accuracy more and more to 98%.