First steps towards autonomous recognition of Monterey Bay ’ s most common midwater organisms : Mining the ROV video database on behalf of the Automated Visual Event Detection ( AVED ) system
msra(2003)
摘要
The development of remotely operated vehicles (ROVs) has revolutionized the world of marine science by allowing quantitative video transects (QVTs) to be recorded underwater. This non-invasive technique allows previously unavailable information to be obtained concerning organism diversity, distribution, and abundance. Unfortunately, processing of these QVTs at the Monterey Bay Aquarium Research Institute (MBARI) is a lengthy and tedious process that prevents rapid data analysis. The development of a neuromorphic computer vision system is currently in progress to solve this problem of the data processing bottleneck. The proposed computer vision system combines a saliency-based attentional module and a recognition module, both designed after the human visual system. The “saliency” model mimics the largely unconscious “bottomup” search mechanism of primates, responding specifically to visual cues in the surrounding environment. The recognition module involves learning of specific targets and requires “top-down” biasing based on learned traits. The attentional module at MBARI can already determine areas of potential interest in the marine environment. However, in order to begin training the computer system for the recognition module, a test database of representative organisms must first be established. The extensive MBARI database was mined to reveal the relative abundance and frequency of annotations of all organisms in the Monterey Bay. Based on the rankings, representative video clips were compiled of the top 25 most common mid-water organisms as a preliminary test database for the recognition module.
更多查看译文
关键词
vars,salience,inhibition-of-return,aved,recognition,bottom-up attention,vims,qvt,winner-take-all,pelagic organisms,rov,video
AI 理解论文
溯源树
样例
生成溯源树,研究论文发展脉络