Скачать 75.03 Kb.
Real-Time Recognition of Automotive Vehicles Using
Advanced Imaging Technology
In today’s world of terrorist threats and amber alerts the ability to easily find and track the movements of specific makes and models of automotive vehicles has become very important. The increasing use of video cameras to monitor traffic by human operators is limited in its scope and performance. If the make, model and year of a vehicle could be automatically recognized from real-time video feeds that information could be used to track movements of suspect vehicles and correlate that movement to a particular situation or scenario.
The proposed research will explore the application of advanced imaging techniques to vehicle recognition. By constraining the views to limit variability it is anticipated that a richer feature set and techniques borrowed from other areas (hand geometry, facial recognition) can be employed.
The scope of the work will be to develop the fundamental techniques and algorithms in a static environment. That is, using still images as if they were extracted from a video feed. Conclusions about image quality parameters required to make a live video feed viable will also be drawn.
Follow on work based on the results is:
CHAPTER 1. INTRODUCTION
Since the development of computers the world is awash in structured, field oriented data. The nature of that data is that it is relatively easy to know where to look for information and the attributes associated with that information. However, the increasing penetration of unstructured data generating technologies (video and audio streams, unstructured textual data on the Internet, etc.) in daily life is creating a new challenge and opportunity. In broad terms, how can these new forms of data be used to advantage? An illustrative example is set in the context of what is known as “Amber Alerts”.
Amber Alerts are used in the case of child abduction. On notification of a child’s abduction, law enforcement agencies broadcast over many different media (am/fm radio, television, and roadside signs) the description of the people involved and usually the make, model and year of the vehicle. Obviously, the general public becomes part of a large, ad-hoc surveillance network.
At the same time, the use of video cameras to monitor and direct traffic flow is becoming common place. With the decreasing cost of video technology, cameras proliferate. With millions of miles of roadway the potential for growth is enormous. However, the feeds from these cameras are monitored by people. The ability to both monitor and track information at any level of detail is constrained by the inherent limits of human interaction.
Identification of Problem and Goal to Be Achieved
Therefore the broad question to be addressed is “Can a new technology be developed or existing technology extended to recognize a vehicle’s make, model, year, color etc, from a live video stream?” In more technical terms, what are the features, recognition algorithms? How can the views of a vehicle from different angles as presented in video streams, be used to strengthen the recognition? Can background clutter be easily eliminated? What is the impact of varying lighting conditions and weather on recognition? What quality of video is required to enable effective recognition? Can the recognition achieved across multiple cameras be coupled to machine learning techniques to achieve an effective unattended system?
Significance of Work
The eventual goal in this research is to create computer methods for automatic identification of objects in a constrained environment. Automated object identification of motor vehicles has many potential uses including and can aid in the following:
Research Questions to Be Investigated
Automatic object detection is a difficult undertaking. The main challenge is the amount of variation in visual appearance. An object detector must cope with the variation within the object category and with the diversity of the visual imagery. For example, cars vary in size, shape, coloring, and other details such as headlights or tires. The lighting, surrounding scenery, distance and angle of the view (an object’s pose) affects its appearance.
The central research issue is how to cope with variation in appearance. The research will investigate how to
Limitations and Delimitations of the Study
The research will develop a model by working with set a of 2 and 4 door passenger cars. Trucks, SUVs, vans, etc. will be excluded. The model will then be applied to vehicles of differing characteristics. The goal will be to generalize the model to the highest degree possible.
CHAPTER 2. RELEVANCE OF THE RESEARCH
In this research we are developing computer and camera methods which will automatically locate automobiles in still photographs. Our goal is to develop algorithms that are accurate and computationally efficient. Our approach is to use statistical modeling to capture the variation in automobile appearance.
Historical Overview of the Theory and Research Literature
Previous approaches used to automatically locate automobiles in still photographs and video include:
However, both these have one fundamental limitation; because of limited computer memory and training data, they use a discrete number of values to describe appearance .
Schneiderman  “Describe a statistical method for 3D object detection. In this method, the 3D geometry of each object is decomposed into a small number of viewpoints. For each viewpoint, a decision rule is constructed that determines if the object is present at that specific orientation. Each decision rule uses the statistics of both object appearance and "non-object" visual appearance. Each set of statistics is represented using a product of histograms. Each histogram represents the joint statistics of a subset of wavelet coefficients and their position on the object. The approach is to use many such histograms representing a wide variety of visual attributes. Using this method, the first algorithm is developed that can reliably detect faces that vary from frontal view to full profile view and the first algorithm that can reliably detect cars over a wide range of viewpoints.”
Rajagopalan, Burlina, Chellappa 
“This paper describes a method for car detection from aerial images. They use a distance based classification metric on 16 x 16 regions. They cluster their training images into several classes of cars and several classes of non-cars. For each 16 x 16 input region, they compute the distance to each class. If the input is closest to a car cluster and under some threshold they classify it as a car. The distance threshold they use could be thought of as a Mahalanobis-like distance metric, except instead of normalizing distance by just 2nd order statistical moments, as in Mahalanobis distance, they use some higher order moments also. They have reported some success in detecting cars from this vantage ”.
Papageorgio, Poggio 
“In this method the Haar wavelet transform is taken of each input region. The wavelet coefficients from two of the middle frequency bands (3,030) wavelet coefficients are used as input to a quadratic classifier. The coefficients in the quadratic classifier are learned by using the Support Vector Machine training method. They report some success in detecting straight-on frontal and straight-on rear views ”.
Schneiderman, Kanade 
“This paper describes a trainable object detector and its instantiations for detecting faces and cars at any size, locations, and poses. To cope with variation in object orientation, the detector uses multiple classifiers, each spanning a different range of orientation. Each of the classifiers determines whether the object is present at a specified size within a fixed-size window. To find the object at any location and size, these classifiers scan the image exhaustively ”.
Summary of What is Known and Unknown
A view based approach works as follows: For each object, several detectors are built where each one is specialized to specific orientation and can accommodate small amounts of variation. To be able to detect an object at any orientation all detectors to the image are applied and the results merged such that they are spatially consistent.
The main unknown is whether an effective mechanism to identify vehicles from a video stream is feasible. Questions such as how many and the nature of the constraints imposed (e.g. which viewpoints to use) are open questions.
Contribution of the Study
The primary contribution of our study will be the creation an approach for the efficient capture / storage / retrieval of objects for the automatic detection of automobiles using still shots extracted from streaming video. A secondary contribution will be laying the groundwork to apply these techniques to detecting automobiles directly from live video feeds, as well as using these techniques to detect other objects.
CHAPTER 3. METHODOLOGY
Research Methods to Be Employed
This research project will use a non-experimental, quantitative research methodology.
The major repository will be a template database of vehicles and their components; second, there will be a set of individual templates of acquired vehicles.
Specific Procedures to Be Employed
The research approach is to set up a camera in various positions around the subject vehicle. Views to be captured will include front, rear, driver side and 45 degree view of driver’s side. The lighting will be held consistent as will the distance from the vehicles. The camera will capture the image of the subject vehicle and pass that image to the feature extractor. Once the features have been extracted, the resulting template will be stored in the template database.
A database of vehicle templates will be used in the matching process. The same camera equipment will be used to acquire the image of the subject vehicle and to acquire the images of the vehicles that will reside in the vehicle template database. This approach to image capture will reduce the likelihood of anomalies attributable to differences in the camera equipment used to acquire the subject and database images.
The feature extraction algorithm will build the templates to be stored in the database and the template of the subject vehicle. A matching algorithm will be used to match the template of the subject vehicle to the templates stored in the database.
The research has been broken down into the following four sub-problems:
Resolution of the images will be varied algorithmically. That is, lower resolutions images will be created from the higher resolution images captured by the cameras.
Once the matches are chosen, the algorithm will rank the candidates by order of score. This is the hybrid approach to identification as described in . The matches and non-matches both will be reviewed to establish the accuracy of the matching algorithm.
We are using the hybrid approach to the matching problem  and the results will be presented as series matrices. Each matrix will report the results of a different variable of our research, such as camera position. The columns of the matrix will indicate the different conditions of the variable, such as where the camera was positioned. The cells of the matrix will contain the percentage of correct matches. See Figure 3.1 for an example of the camera position matrix.
Fig. 3.1 – Results: Camera Position
We anticipate that this research will influence several different communities including law enforcement and national security. We feel the most realistic outcomes are:
Personnel requirements include one Project Director, one Testing Coordinator, two Researchers, two Software Developers, and one Statistician. The Project Director and Testing Coordinator will work on the project full-time at 100% salary and all other staff members will work part-time at 50% of salary.
Hardware requirements include camera equipment, PCs, servers, printers, and scanners. Facilities include office space for full-time and part-time personnel; storage of camera equipment and setup of computer equipment. In addition, office furniture & fixtures and office supplies will be purchased.
Reliability and Validity
Data that is established from research must be shown to be both reliable and valid. Stability reliability of the data from this research will be demonstrated through the test-retest approach . We fully expect our results to be repeatable through multiple repetitions.
Criterion-related validity is classified as predictive validity, concurrent validity, convergent validity, and discriminate validity . Predictive validity is specifically defined in  as “the operationalization's ability to predict something it should theoretically be able to predict.” We think that predictive validity, by definition, is the most appropriate measure to use for our research.
CHAPTER 4. DISCUSSION OF EXPECTATIONS
In the research, feature extraction techniques are studied. Such features can be categorized as coarse to fine features. Coarse features such as side view of a vehicle, size of a vehicle, may be used to identify or reject a certain type of vehicle such as truck or car very efficiently in real time. The fine features such as position of front light, grill size, windshield may be used to identify a certain model of the vehicle. These identified features are not only useful for vehicle identification but may also be utilized as indices to vehicle feature database for fast search. In many cases, the search of the database depends on the details of the vehicle that are provided. The processes of identifying a white truck or a 1999 white Ford Mustang convertible involve different levels of search. It should be expected that the system should take less time to match a vehicle to less specific (or coarse) feature description.
Also, experimental images of each sampled vehicle will be captured from multiple views. It would then be possible to determine which view of the vehicle is most effective in vehicles identification. The information is important for determining the optimal position of the camera. In a common camera surveillance situation, a limited number of cameras (usually one or two cameras) are used to scan a large area. For example, in the case that a camera is used to cover a multi-lane highway, each scanned picture may contain a number of vehicles and the images of each vehicle may be at slightly different angles.
The research will explore the application of advanced imaging techniques to vehicle recognition and in doing so advance the state of art in 3D object recognition and identification using real-time video feeds. The research will focus on the detection of objects in a static environment with the intention to adapt these techniques to the dynamic environment of live video feeds outside the scope of this proposal. The outcome of the research will be the following:
Practical Applications of the Findings
The main application of this research is real time vehicle identification system.
The identification process must be near real time to be useful for the purpose. If this can be accomplished then automatic object detection and recognition can be used to extract more information from images and help automatically label and categorize them. By making databases of large digital image collections easier to search, they will become accessible to wider groups of users and greatly enhance efforts such as the Amber Alert project allowing precious time to be saved in locating suspicious vehicles that match a set of features.
Constraints and Limitations of the Study
Object detection is difficult because images contain a large amount of data. Computer power and memory available for the study limit the size of digital image collection used which may cause some variation to occur between the physical world and study results.
In our study, the same camera equipment will be used to acquire the image of the subject vehicle and to acquire the images of the vehicles that will reside in the vehicle template database. In real world situation, it is possible that different image capture devices will be used in both the enrolling and identifying stage. The amount of anomalies attributable to differences in the camera equipment used to acquire the subject and database images are out of the scope of the study.
Recommendations for Additional Studies
There are several research problems that we see as a natural continuation of the work:
Contributions to the Field and Advancement of Knowledge
The proposed research will advance the state of the art in 3D object detection in the following ways:
REFERENCES and LITERATURE REVIEW
Данная курсовая работа была посвящена исследованию и реализации Real time keying'a (кеинг в реальном времени). Проведены работы по...
Россия, Ярославль, ул. Советская, 14. Тел. (4852) 79-77-75. E-mail