Off-campus SJSU users: To download campus access projects, please use the following link to log into our proxy server with your SJSU library user name and PIN.
Master's Project - Campus Access Only
The lessons evolution has ingrained in us is the need to see, perceive and engage with our environment. Computer vision is inspired from the same need to develop machines into a new era of technology where they can utilize their computing power to see, understand and make decisions on their own. In this project, we look at how the computer vision field has grown over the years with the power of artificial neural network at its core. We explore latest research for motion and gesture detection employing Convolutional neural networks to perform spatial feature extraction from images. We bolster the need of recurrent neural networks like Long Short-Term Memory in conjunction with Convolutional neural networks for learning sequential data and the limitations in achieving real-time accurate visual performance system. We look at how to capture spatial and temporal features using CNN coupled with LSTM. We observed that in case of sparse data we need to be extra careful about overfitting. Sparse data in our images is handled by tweaking the filter size of CNN. We explore 2 approaches of connecting CNN with LSTM. Concluding we see that although 3D CNN’s work extremely well compared to 2D CNN over the dataset. 3D CNN’s don’t generalize well in real time relative to 2D CNN.
Khandelwal, Kushal, "Gesture Recognition with Deep Learning" (2018). Master's Projects. 645.