| Summary: | 博士 === 中央警察大學 === 鑑識科學研究所 === 100 === Abstract
Databases have been widely applied to forensic science and crime investigation.
In forensic science, we can identify questioned evidences or find similar evidences
with these database systems, such as fingerprints, DNA, shoe prints, faces, drug
tablets, video data. In crime investigation, we can use these systems to find the
relationship between different cases.
The traditional user interface of image databases uses textbased
retrieval
techniques which use text tags to label images. However, textbased
image retrieval
systems require manual labeling which is a cumbersome and expensive task for large
image databases. Furthermore, the variety of keywords from diversity realization may
cause different retrieval results.
At the beginning of the 1980s, digitized fingerprint databases became the first
forensic databases to be widely used. Other image databases, such as shoe marks, tool
marks and striation marks on cartridge cases and bullets, also became popular. The
improvements of image acquisition and storage facilities make it economically
feasible to build color image databases. With automatic contentbased
comparison
algorithms, we can find similar images from databases. The development of a
retrieval system requires a multidisciplinary approach with knowledge of multimedia
database organization, pattern recognition, image analysis and user interfaces. The
most important knowledge is contentbased
image retrieval (CBIR) techniques that
have been subjected to intensive research efforts.
CBIR provides a good tool to retrieve interested images from image databases. It
use “image features” (instead of “text”) as “searching keywords”. Some commercial CBIR have been available, such as Integrated Automated Fingerprint Identification
System (IAFIS), Integrated Ballistic Identification System (IBIS), TreadMark™ for
shoe prints, and Forensic Information System for Handwriting (FISH).
Video data, especially from surveillance systems, also play an important role in
forensic science and crime investigation. There are a lot of digital video data collected
in database nowadays. Video analyzing technologies are useful for us to quickly
access and get information from those video data. Motion detection is one of useful
video analyzing technologies. Motion detection is used to segment interested image
areas and find possible moving objects in video data. In general, motion detection is a
process of confirming a change between a moving object and its surroundings or the
change in the surroundings relative to an object. However, this process is sensitive to
the light condition. In this thesis, we propose an efficient motion detection method for
false alarms.
Automatic license plate recognition system (ALPRS) is one of the most
important examples of applying computer techniques to intelligent surveillance
systems. ALPRS has been applied in three main categories: (1) Road traffic
management: improving the flow and safety of vehicle traffic controls. (2) Security
management: recognizing and controlling the conditions of entry and exit of vehicles
from parking areas and restrained regions; tracking of vehicles in the restrained region.
(3) Crime prevention: help reducing the criminal intention before crime incidents
happen; help investigating and tracking criminal vehicle(s).
The performance of the license plate localization is crucial to ALPRS, because it
directly influences the accuracy and efficiency of the plate number recognition. A
number of methods have been proposed for license plate location. Most literatures focus on detecting the accurate location of single license plate from a vehicle image or
video. However, there are usually more than one vehicle appear within an image
frame simultaneously in practical cases. That is, we need to locate multiple license
plates before identifying their license plate numbers. In this thesis, we use the optical
flow algorithm and blob analysis to locate multiple license plates in video sequences.
The main contributions of this thesis are as follows: (1) provide the edge
orientation features to retrieve crime scene images, shoe print blocks, and geometric
objects; (2) design a feasible image database of drug tablets based upon shape
signatures and texture features; (3) provide a motion estimation method to detect the
motion incident of false alarms and locate multiple vehicle license plates. The
experimental results show the capability of the proposed systems and methods.
|
|---|
