Copyright © 2014 by Intelligent Systems Laboratory, Computer Science Department, Technion - Israel Institute of Technology, Haifa 3200003, Israel. All rights reserved

Far Field Surveillance Target Classification

Far Field Surveillance Target ClassificationBy: Amir Geva

Automated processing of surveillance video is a growing field. Classification of targets in surveillance videos is of great importance for efficient security. This thesis report summarizes research on the problem of classification of far field targets in surveillance videos. Far field targets are too far, and thus too small for standard classification methods. Classification of very small targets requires the utilization of all information available in the video sequence. The research covered various parts involved in the construction of a robust and accurate classifier. It is assumed that the classifier is a part of a larger system that includes object detection, tracking, and segmentation which are not in scope of this research. The re-port describes novelties in the steps of: feature extraction, covering static shape features and also introducing motion based features; feature selection, describing a genetic algorithm wrapper method to select the optimal subset of features; sequence classification which takes information from a sequence of multiple frames in order to make a decision that is impossible based on a single frame using a method of error correction output codes and delayed voting; and occlusion handling, which filters out instances of the target that are not fully visible. The features selected bring a simple voting classifier to a level of 91.4% accuracy, whereas the novelties that take advantage of the sequences and filter out noise, as described in our research, show an improvement that brings the resulting classifier to an accuracy level of over 95.7% on the test benchmark.


Read more ...

Estimating Pose and Motion using Bundle Adjustment and Digital Elevation Model Constraints

Estimating Pose and Motion using Bundle Adjustment and Digital Elevation Model ConstraintsBy: Gil Briskin

In spite of the e ort placed in nding ecient and robust estimates for the pose and motion of a calibrated camera from multiple image views, the problem continues to attract extensive attention in the photogrammetric and computer vision communities. Perhaps the main reason for this continuous attention is the well-known fact that pose and motion cannot be uniquely solved from a series of images. Some of the limitations are obvious from the start: one cannot expect to obtain absolute information about pose with respect to an external coordinate system from a sequence of images; other limitations are more subtle and relate to the speci cs of the visual information. One such limitation is the pose estimation when the images are taken from a small baseline relatively to the scene. The lack of uniqueness makes the problem hard and ill-conditioned, and hence additional assumptions, external information and specially design algorithm are required to produce robust and reliable estimates. This thesis will present a new approach for computing the pose and motion problem based on adding external absolute information, namely, information with respect to the external coordinate systems, to the data obtained from the sequence of images. 

Read more ...

Estimating Epipolar Geometry for Mobile Robots

Estimating Epipolar Geometry for Mobile RobotsBy: Cherevatsky Boris

This thesis addresses the problem of robot navigation using natural visual features in a planar environment. Suppose we have a mobile robot equipped with a camera. The robot can move on the oor and capture images with the camera, and save them on a storage. The robot acts inside a room which is surrounded by walls, which are orthogonal to the floor. Many visual features appear on the walls, and the robots goal is to go from one pose to another by using only those visual features. 


Read more ...

Dimensionality Reduction for 3D Articulated Body Tracking and Human Action Analusis

Dimensionality Reduction for 3D Articulated Body Tracking and Human Action AnalusisBy: Leonid Raskin

Tracking humans, understanding their actions and interpreting them are crucial to a great variety of applications. Tracking is used in automated surveillance, human-computer interface applications and in security applications.






Read more ...

CFORB Visual Odometry Project

CFORB Visual Odometry ProjectBy: Daniel Mankowitz

This report details the CFORB algorithm. This algorithm utilizes an ORB detector [7] to detect features and a FREAK descriptor [30] in order to compute feature descriptors. Matching features between images is computed using the Hamming distance. This algorithm is invariant to changes in both scale and rotation. In addition, two geometric constraints have been utilized in this algorithm in order to remove bad matches between images. This algorithm has been tested on both the KITTI outdoor dataset and the Tsukuba indoor dataset. The algorithm achieves an average translational error of 3:73% on the KITTI dataset which places it among the top 25 algorithms for this benchmark dataset. It also has a rotational error of 0:0107deg=m. An average translational error of 2:4% has been achieved on subsets of the KITTI dataset where there is a large amount of texture and uniform feature spreads across images. CFORB also performs well on the Tsukuba dataset and achieves an average translational error of 3:70%. Finally, an overview of state-of-the-art Visual Odometry algorithms is presented followed by a description of the Visual Odometry problem and feature detection techniques in the Appendix.


Read more ...