Project Motivation and Description:
The topic of my senior design project at Johns Hopkins University was detecting animals with action recognition without a camera, a form of acoustic surveillance system. Acoustic surveillance systems offer many advantages over electromagnetic radars because of lower cost, simpler signal processing, and immunity from electromagnetic interference. For Zhaonian Zhang’s dissertation in 2008, he designed an active acoustic sensing system prototype for imaging moving objects such as a human walking or deer running. Without using a camera, the prototype works via transmitting an ultrasonic waves which hits a moving object, disrupts the waves, and a receives back the disrupted waves. The system knows that there is moving object in front of it by this disruption, which is the micro-Doppler effect. The goal was to upgrade this prototype for 2016 usage as well as implement a machine learning algorithm to classify an animal from its ultrasonic signal.
Team and Role:
I served as the product manager of this project. Specifically, I miniaturized the prototype via removing extraneous parts–resistors, capacitors, and the extra transmitter, added in a new microcontroller–a PIC board, wrote the assembly code for the new microcontroller, and added in a wireless capability. My teammates handled 3-d printing a new case to house the new design, systems engineering/cost analysis, and coded up a conditional random field (CRF) clustering model, a semi-supervised learning technique, using MATLAB to accurately identify the animals based on their respective signal.
Most challenging part of the project:
Zhang’s original schema of his prototype and his parts list was not available , so I had to re-create the schema using the PCB Artist software . I would not necessarily label this as the most challenging, but more time-consuming because most of the parts used on PCB Artists were too new for the older board, so I lost a week designing the older parts to be used in the schema. Although writing the assembly code was tricky at first, I found online documentation to help.
Moving Forward:
Although this project is over, if I had more time I would have wanted to use a different microcontroller (potentially a ARM processor) that better supports machine learning algorithms. Although a PIC controller is not terrible, an ARM processor would have only helped. In addition, though I only had time to run one test after my teammates wrote the CRF clustering code, I wish I could have ran more tests to truly measure the accuracy of the algorithm.