Researchers place in video analytics competition
An ADSC research team ranked in the top 3 in a recent Tech Challenge, hosted by the Infocomm Development Authority of Singapore (IDA). The competition focused on video analytics and asked teams to develop a solution to anonymize videos being used for crowd sensing and traffic monitoring in real time.
“Detecting humans in a crowded scene has been a very challenging computer vision task for a long time,” ADSC Adjunct Research Scientist Bingbing Ni said. “The current performance is still not mature. We wanted to create a novel human detection system that has high accuracy and high reliability and is close to real deployment.”
The competition included two stages where participants created algorithms to detect and blur out human faces appearing in public surveillance videos around Singapore. The videos in stage two were more challenging to anonymize, as there were many more people and partially hidden faces compared to stage one. IDA hosted the challenge to encourage Singapore researchers to discover improved video analytics techniques in order to gain more meaningful insights from video data.
“One of the difficulties of this challenge was that the quality of the videos was so terrible,” ADSC Software Engineer Tan-Loc Truong said. “The color, illumination condition and the number of people fluctuated in an unpredictable way. This meant that in order to be able to detect all the faces, our algorithm had to be robust enough to address all of these conditions.”
ADSC’s team also included Research Scientist Jiwen Lu, former intern Duyen Phuc Nguyen and Illinois Electrical and Computer Engineering Professor Pierre Moulin. The group’s research at ADSC focuses on designing architecture for multisensory, multimodal analytics and applying it to audiovisual surveillance and assisted living. The algorithms they use to detect rigid objects in digital images were useful in helping develop the algorithms for the IDA Challenge.
Participants’ algorithms were evaluated on their ability to identify objects in the video, how accurate their algorithm was and how fast the algorithm performed. ADSC’s team was second in recall and first in precision and timing in stage one and third overall in stage 2. The team received S$3,000 for their stage one 2nd place finish.