Background-Centric Approach for Moving Object Detection in Moving Cameras

Abstract

A number of surveillance cameras have been installed for safety and security in actual environments. To achieve a human-level visual intelligence via cameras, there has been much effort to develop many computer vision algorithms realizing the various visual functions from low level to high level. Among them, the moving object detection is a fundamental function because the attention to a moving object is essential to understand its high-level behavior. Most of moving object detection algorithms in a fixed camera adopt the background-centric modeling approach. However, the background-centric approach does not work well in a moving camera because the modeling of moving background in an online way is challengeable. Until now, most algorithms for the object detection in a moving camera have relied on the object-centric approach using appearance-based recognition schemes. However, the object-centric approach suffers from the heavy computational complexity. In this thesis, we propose an efficient and robust scheme based on the background-centric approach to detect moving objects in the dynamic background environments using moving cameras. To tackle the challenges arising from the dynamic background, in this thesis, we deal with four problems: false positives from inaccurate camera motion estimation, sudden scene changes such as illumination, slow moving object relative to camera movement, and motion model limitation in a dashcam video.

To solve the false positives due to motion estimation error, we propose a new scheme to improve the robustness of moving object detection in a moving camera. To lessen the influence of background motion, we adopt a dual-mode kernel model that builds two background models using a grid-based modeling. In addition, to reduce the false detections and the missing of true objects, we introduce an attentional sampling scheme based on spatio-temporal properties of moving objects. From the spatio-temporal properties, we build a foreground probability map and generate a sampling map which selects the candidate pixels to find the actual objects. We apply the background subtraction and model update with attention to only the selected pixels.

To resolve sudden scene changes and slow moving object problems, we propose a situation-aware background learning method that handles dynamic scenes for moving object detection in a moving camera. We suggest new modules that utilizes situation variables and builds a background model adaptively. Our method compensates for camera movement and updates the background model according to the situation variables. The situation-aware scheme enables the algorithm to build a clear background model without contamination by the foreground.

To overcome the limitation of motion model in a dashcam video, we propose a prior-based attentional update scheme to handle dynamic scene changes. Motivated by the center-focused and structure-focused tendencies of human attention, we extend the compensation-based method that focuses on the center changes and neglects minor changes on the important scene structure. The center-focused tendency is implemented by increasing the learning rate of the boundary region through the multiplication of the attention map and the age model. The structure-focused tendency is used to build a robust background model through the model selection after the road and sky region are estimated.

In experiments, the proposed framework shows its efficiency and robustness through qualitative and quantitative comparison evaluation with the state-of-the arts. Through the first scheme, it takes only 4.8 ms in one frame processing without parallel processing. The second scheme enables to adapt rapidly changing scenes while maintaining the performance and speed. Through the third scheme for the driving situation, successful results are shown in background modeling and moving object detection in dashcam videos.