Stereo Camera vs 3D Laser Scanner vs SDP Fusion Algorithm

Depth perception is a fundamental problem in computer vision with many applications such as robot navigation and 3D video. Significant progress has been made on depth sensing using passive (e.g., stereo) and active (e.g., laser, TOF) sensing. However, inaccurate results of stereo imaging caused by the ambiguity problem in stereo matching, and the slow, noisy and sparse results of laser scanners and time-of-flight (TOF) cameras fail to meet the requirements of many applications. In recent years, fusion of active and passive sensing has shown potential for addressing this problem. Our work proposes a super-resolution algorithm using a bilateral filter to upsample small depth images from an active sensor, with the help of a high-resolution camera image. The proposed method obtains considerably more accurate results (an error reduction of up to 6.8x) compared to the literature, with sharper and more realistic edge definition. This is shown quantitatively using both the Middlebury benchmark and a set of laser-scanned scenes. The proposed method is better suited to the sparse and unevenly-distributed depth measurements from active depth sensors, and it requires minimal parameter adjustment across different scenes, which makes it practical for real-world applications.

This video compares 3D images obtained by:
- An off-the-shelf stereo camera (Point Grey Bumblebee XB3)
- A 3D laser scanner (SICK LMS 291)
- Our proposed laser-camera fusion algorithm (SDP)

For more information, visit: http://ali.kashani.ca/active-passive-fusion

Category:

Education

Tags:

• laser scanner
• camera
• upsampling
• super-resolution
• range image
• 3d
• sensor fusion
• integration
• ubc main library
• outdoor environment