As a part of project, I experimented with the different network models and parameters of Neural Networks

useful for classification. The project was implemented using Matlab and Visual C++

Download an Informal Report Here

The project involves recognizing gestures made using the SIMPUTER.The Simputer is a hand held device that runs on a customized version of Linux. It is a powerful computational tool and has won much applause in both the academia and industry. The SIMPUTER is the brain child of Prof..V.Vinay,CSA, Indian Institute of Science and  Chairman String Labs and  Prof. Swami Manohar, CSA, Indian Institute of Science, CEO String Labs (Picopeta Simputers Pvt Ltd).

The SIMPUTER is equipped with a 2D accelerometer that can measure accelerations between -0.2g to +0.2g. The accelerations are sufficient data for us to reconstruct patterns and recognize gestures. 

The project is divided into two phases

1) 'The Clustering phase' using a semantic approach 

2) 'The Pattern Recognition phase'

So far, in the Clustering phase I have experimented with three types of algorithms. 

(i) Employing the K-Means clustering algorithm I came up with K clusters that will uniquely define each gesture. But there were inconsistencies due to the initial random distribution of the clusters. 

(ii) In the second method, I employed a dynamic size sliding window algorithm. Then taking the means of each window, I created clusters applying the First Order Markov Chain principles. Thus, I propose that the system is in a particular state I at time T=q, only if it was at state J at time T=q-1. This algorithm gave much improved performance in providing state information.

(iii) In this method which I am currently working on, I hope to use a powerful mathematical tool called Singular Valued Decomposition or SVDs to reduce the information present in a loss less and efficient manner. SVDs have been widely employed in genetics and other related fields to capture information effectively and efficiently. Prof V Vinay is an authority in this field and has been guiding me constantly.

In the Pattern Recognition Phase, I have so far experimented with Multi Layer Percpetrons (MLP) with Back Propagation (BP). I have used this model with the first two clustering techniques. I was able to generate a low error tolerant system that produces an accuracy of 95% -gesture recognition using the sliding window algorithm for clustering. 

In the next phases of the project I hope to develop a robust mathematical model for gesture recognition using Hidden Markov Models. I have chosen this model based on its remarkable success in speech recognition. From my preliminary studies I feel that it is possible to liken a constrained form of gesture recognition to speech recognition. Hidden Markov Models are useful dynamic gesture recognition and hence it is a model, I feel, worth experimenting with.

www.amidasimputer.com

www.picopeta.com