Facilitating PTZ Camera Auto-Calibration to Be Noise Resilient With Two Images

Abstract

This paper presents a simple but effective technique to calibrate a PTZ (Pan/Tilt/Zoom) camera by using only two images for five intrinsic parameters: focal length, aspect ratio, the principle point coordinates and the distortion coefficient. In our approach, the SCC-SURF (Shape-Color Combined SURF) descriptor is first employed to obtain robust point correspondences in a pair of color images taken before and after the camera undergoing an arbitrary pan-tilt rotation respectively. Based on the radial lens distortion division model, the point correspondences between these two images are applied to calculate the homography and the distortion coefficient simultaneously. The estimated homography is proved more precise with our novel framework CWRLD (Covariance Weighted Ransac under Lens Distortion), which employs a covariance matrix in the presence of feature location noise. Finally, the remaining four intrinsic parameters are solved using directly decomposing estimated homography with a series of Givens rotations. Both synthetic and real data are provided to verify that our proposed technique is precise, convenient, and applicable for online calibration without regard for a specific imaged environment.

SFA-MDEN: Semantic-Feature-Aided Monocular Depth Estimation Network Using Dual Branches

“…The unsupervised methods based on stereo image pairs simulate the stereo vision system, utilizing neural network learning, and create a virtual binocular system [ 35 ] through monocular images, while the other type of unsupervised method based on image sequence simulate the camera movement through multi-frame images. Zhou et al [ 36 ] proposed an end-to-end approach, where the supervision signal comes from the left and right adjacent frames, which are warped to the target frame based on the perspective projection model, and the training cost is computed by comparing the warped frame with the real target frame.…”