4-D Distributed Modeling and Visualization
Technical Strategies
We address the problem of acquiring, fusing, manipulating, and visualizing multiple image/video/data information streams from a group (possibly thousands) of robot or manned sensors. Image sensor can be on the ground as head-worn cameras, robotic panoramic cameras, aerial camera platforms, etc. They can be fixed or moving and their data can be arriving live in real-time, or coming from an archive. The problem we really address is providing some way for a person or group of people to digest and understand all this information easily with some sense of the spatial, temporal, and truthful nature of the information. The research fronts are stated as followings towards to this effort (Figure 1):
Data acquiring: our work on hybrid
sensor packages for outdoor tracking has progressed to the point where we have
a differential GPS system (base station and mobile) operational around our area
of the USC campus. We are adding compass and gyro sensors to complete an
initial system that will allow us to move freely (relatively) around the campus
gathering image, data, and video. We expect to display that data on the LiDAR
base model we have.
Data
fusion and integration: trivially,
a wall screen can be filled with video or still images, abstract data, maps,
and text communications. The cognitive load this presents to a commander or
strategist is however overwhelming. Information overload is only likely to get
worse as more computing, sensing, and communications pervade the battlefield.
The problem of making all this information digestible and understandable
depends on having a coherent presentation and fusion framework that allows a
user to easily understand relationships and switch focus between levels of
detail and specific spatial or temporal aspects of the data.
We feel that two elements are central to making this kind of information presentation possible:
1) detailed geometric models of the environment viewed freely over a range of scales and directions
2) visual images and abstractions of information displayed in the context of the scene geometry
Tracking and mode refinement: underlying both of these is a need for tracking of the sensors. Only by knowing where the sensors are and where they’re acquiring data can their information be properly placed on the model. We consider the case where images (or other data) from possibly hundreds of robots, aerial platforms, or moving personnel arrive back at a command post. If their tracked positions/orientations are known, these images can all be projected onto the scene model, thereby presenting the observer with a single coherent and evolving view or the complete scene.
Our tracking efforts focused on pose from lines and auto-calibration of lines. We feel that since lines or edges are dominant features of man-made strictures, we should make full use of them for tracking. We expect to both use lines for computing/refining camera pose and refining structure edges based on auto-calibration of line/edge features.
Dynamic texture projection: we address the image-texture projection problem with the twist that we also consider the correction of the tracking of the sensor and the refinement of the models from the same image data. The availability and low cost of LiDAR data is something we leverage. We are developing methods to segment buildings and plant growth. Since LiDAR comes in unorganized point clouds, we also need to develop or use tessellation methods.
[ MURI Home Page | Executive Summary | Our Mission | Technical Strategies ]
[Results | Presentation | Technical Staff | Related Link | Event and News]
Question to Suya You