Gesture Recognition System Using Particle Filter

We will be presenting a gesture recognition system based on particle filters. The accelerometer data from the UTD MHAD (Multi-modal Human Action Dataset) has been used to create the gesture templates using Dynamic Time Warping Barycenter Averaging. The input gesture is then matched with the template gestures to determine its belongingness to a class depending on the probability of the particles associated with each gesture. We manage to achieve an average of 87% accuracy for gestures in UTD MHAD and 82% average for UC Berkeley MHAD. The low-power hardware architecture for this model has also been designed in Quartus and checked in Modelsim for validation, it manages to achieve the accuracy with marginal 2% error in bit conversion.