MU Digital Repository
Logo

Robust object detection and Tracking using stereo vision

Raghavendra, U (2014) Robust object detection and Tracking using stereo vision. Phd. Thesis thesis, Manipal Institute of Technology, Manipal.

[img] PDF
Raghavendra U.pdf - Submitted Version
Restricted to Registered users only

Download (77MB) | Request a copy

Abstract

Video Surveillance for multiple object is a challenging areas of research in computer vision. It comprises of two correlated components; object detection and object tracking which are key requirements in variety of intelligent applications namely, early recognition of on-going abnormal activities, automated surveillance, mobile robot navigation, advanced driver assistance system, etc. Object detection is to identify the location of object in a scene and tracking is to associate the detected object along a sequence of frames 2D computer vision techniques and the stereo vision based 3D techniques are the two major approaches developed in the last few decades for object detection and tracking. Majority of the these techniques are designed under completely controlled environment and are prone to ine�ciency due to real-time di�culties such as objects count, illumination variations, object orientation, pose, occlusion etc. Even though few stereo vision based 3D techniques have made an attempt to address these di�culties, they failed to achieve good accuracy. Consequently, a solution to the above problems within a single framework is the vital requirement of the present system. This research work developed robust object detection and tracking method using stereo vision technique. The major contribution of this research lies in three modules of stereo vision based object detection and tracking namely, disparity estimation, object detection and object tracking In the �rst module, stereo matching algorithms are developed for disparity estimation in order to address the real-time di�culties such as depth discontinuity and radiometric variations. Initially, it developed a stereo matching algorithm using anchor diagonal based shape adaptive local support region to handle depth discontinuities in the stereo image. Developed algorithm is simple in construction and outperforms many local and global stereo algorithms. In addition, two di�erent stereo energy models are developed for compensating radiometric variations between stereo image. First, a Hybrid Correlation (HC) based stereo energy model is developed for disparity estimation using two e�cient correlation measures. Developed HC has been extended to introduce a robust stereo energy model known as Correlation Fusion (CF). Developed disparity estimation algorithms are evaluated using benchmarked stereo data sets with di�erent illumination and exposure conditions and the obtained results show that it outperforms conventional and state-of-the-art techniques. 3D world Euclidean co-ordinate points of the estimated disparities are computed using triangulation and a 3D Region of Interest (ROI) is generated from it. These 3D ROI disparity points are further re�ned and projected onto the groundplane to construct a new plan-view statistics known as Signi�cance map (Smap). This Signi�cance map can handle multiple object occlusion and uses human biometric information to overcome under and over segmentation problems. Observed features on the Signi�cance map are clustered using connected component based clustering and the position of the object is detected using depth layering technique. Developed detection algorithm is evaluated using a set of test stereo video sequences and the evaluation results are compared with the competing method. Finally, Kalman �lter based multiple object tracking algorithm is developed under continuous detect-and-track framework. It includes an object detection rollback loop with the tracking algorithm, which enhances both detection and tracking accuracies The developed stereo vision based object detection and tracking method has made an attempt to minimize several primary constraining assumptions such as illumination variation, object occlusion, shadow, etc., in a single framework. It has also minimized some of the secondary constraining assumptions such as object pose, camera height, orientations, rotations, etc. The rigorous evaluation of developed algorithm evidenced its robustness for many real-time conditions. Developed system has opened numerous aspects of potential investigation, which can be drawn-out in numerous ways for further direction of research

Item Type: Thesis (Phd. Thesis)
Subjects: Engineering > MIT Manipal > Computer Science and Engineering
Depositing User: MIT Library
Date Deposited: 03 Dec 2014 07:40
Last Modified: 03 Dec 2014 07:40
URI: http://eprints.manipal.edu/id/eprint/140813

Actions (login required)

View Item View Item