Hey UW Bothell! It’s been quite some time since I have been able to post anything. Over the past two months I have been extremely busy finishing up my research and wrapping up my studies here at Osaka University. Now I finally have some time to relax, and with just a week left in Japan I figured I would write about what I’ve been up to.
Over the past five months I have been working in the Department of Intelligent Media at Osaka University, which specializes in computer vision. The key focus of my laboratory is gait recognition, which is a biometric identification method used to identify people based on the way they walk. We each walk in our own unique way, and much like our fingerprints and irises, our gait is unique enough to be used to identify us. Gait identification is especially useful in the field of law enforcement because it is possible to identify someone’s gait regardless of whether that person is cooperative. For example, a criminal might hide their face or wear gloves to cover their fingerprints, but there is no real way for them to cover up their gait.
While most of my lab mates focused on human gait recognition, I was assigned the peculiar task of researching cow gait recognition. You might be asking, why would we need to identify cows? Well, an accurate means of identifying cows could lead to fully automated beef and dairy farms. With beef one of the largest enterprises in the agricultural sector, successful automation could completely change the industry and benefit both farmers and consumers alike.
Unfortunately, current gait recognition techniques achieve unsatisfactory success rates when applied to cows. Technologies like Gait Energy Image (GEI) comparison achieve a success rate of just 60% when attempted on cows, which is nowhere near enough for any sort of automation. One possible reason these methods did not work as well on cows as they did on humans, was because of some of the features of a cow’s walk cycle. Cows tend to move their heads and tails a lot during their gait, which we felt wasn’t particularly useful for identification.
Here is a picture of the difference between two GEIs for the same cow, which illustrates the problems that the cow’s head can cause for gait identification. The red areas are areas that do not match, and as you can see there is a lot of red around the head of the cow despite the two GEIs belonging to the same cow. This difference in the head area could lead to a false identification if not handled correctly.
Based on this, my job was to develop an algorithm which improved identification success rates by eliminating the effects of the motion in a cow’s head and tail. I won’t bore you with too many details, but my algorithm involved splitting the GEIs into subsections and weighing each section based on a probability density function which was retrieved during training. This allowed us to dampen the effects of unhelpful areas of the GEI while enforcing the effects of more helpful areas, thus improving success rates. In the end I was able to improve success rates by about 13% and learn a bunch of things about a cow’s gait that we had not considered. For example, it seems like the torso of the cow would be an excellent area for identification, but our tests showed that this area was in fact the worst area in terms of success rates.
A couple of weeks ago I presented my research alongside other study abroad students in the Frontier Lab program, and I submitted my final report just a few days ago. All that’s left now is to document my final code for any future students who are interested in continuing this work. The experience as a whole was a ton of work, but I learned a lot about working in a research environment, and I was able to pick up some great techniques for various aspects of research which will no doubt help me in my future career.
Thanks for reading!