Ph.D. Proposal
Detecting Objects with High Accuracy and in Real-time: A Vision-based
Scene-specific Object Detector in Mobile Systems with Human-in-the-loop Training
Zheng Li
10:00am 15 September 2015, ITE 325b
In computer vision, researchers pursue to train machines to detect objects as well as humans — with high accuracy and in real-time. Though the goal of highly intelligent machine vision has been the target of research for years, machines still perform inferior to humans. Present research continues to specifically investigate new robust features types that lead to improvement of effective detection accuracy. While use of carefully hand-engineered features usually helps, it requires decades of expertise effort to design a good feature representation. Moreover, the machine-end real-time performance often suffers due to the complicated feature extraction and matching. In application where low latency is as critical as high accuracy, such as with unmanned aerial vehicles (UAVs), or assistive guidance and navigation systems for people with visual impairments, approaches to achieve lower execution times are required.
In this proposal, a vision-based Scene-Specific object Detector (SSD) is proposed which transforms the general vision problem into scene-specific sub-problems in order to incorporate scene-specific a priori knowledge to achieve higher detection accuracy and real-time performance. This SSD deeply involves human-in-the-loop training to acquire possible a priori knowledge. With the combination of human-acquired a priori information and sensed real-time information from multi-sensors, a hierarchical coarse-grain to fine-grain search scheme can be used to detect objects efficiently and robustly in a real-time hardware platform. Such a solution can achieve performance exceeding traditional state-of-the-art approaches.
Committee: Drs. Ryan Robucci (chair), Nilanjan Banerjee, Chein-I Chang, Ting Zhu