Hee Jin

Publication Date


Document Type

Honors Thesis


Computer Science


Neural networks (Computer science), Convolutions (Mathematics), Wireless sensor networks, Image processing, Fourier transformations, Convolutional neural network, Fourier transform, Neural network


Visual Sensor Network is a part of Wireless Sensor Network that uses camera nodes to acquire visual information of the surroundings. Images from these sensors provide useful information for tracking purposes even though there are challenges due to the large data size. Many targets tracking algorithms have been proposed that minimizes these problems and effectively track the target, but not much of this research has been applied to multiple target tracking especially not with Convolutional Neural Networks on microprocessors that could potentially serve as a node in a wireless sensor network. An algorithm is therefore proposed with optimization features such as the Fast Fourier Transform calculations and Overlap-and-Save methods. With these algorithms implemented, training and testing has been done with the online Char74k dataset and number written on paper were used to test the result on the Raspberry Pi node. The training process output training accuracy of 91.2 % with binary outputs and 86 % with 10 outputs. The testing performed on the Raspberry nodes had similar accuracies with an average of 8 correct predictions out of every 10 inputs. For 10 outputs, the accuracy was 6.5 correct out of each 10 inputs




vi, 58 pages : illustrations (some color). Honors project, Smith College, 2016. Includes bibliographical references (pages 42-44)