New machine vision approach to vehicle identification
By Mark Williamson, Firstsight Vision, Tongham, UK; www.firstsightvision.co.uk Machine-vision specialist Firstsight Vision has applied a well-established image-processing method for the first time to real-time vehicle recognition applications.
By Mark Williamson, Firstsight Vision, Tongham, UK; www.firstsightvision.co.uk
MARK WILLIAMSON is a director of Firstsight Vision and chairman of the UK Industrial Vision Association; email@example.com
Machine-vision specialist Firstsight Vision (Tongham, UK; www.firstsightvision.co.uk) has applied a well-established image-processing method for the first time to real-time vehicle recognition applications. In the first application, the class of vehicle (motor-cycle, car, taxi, light van, bus, truck, and so forth) is identified, while in the second, the vehicle manufacturer (Ford, Mercedes, BMW) can be determined. Both methods can be linked in with license-plate reading systems.
The analysis is based on an object-recognition tool that uses statistical learning theories. Promising results have been obtained in some preliminary trials, and the company now needs to work with a dedicated partner to ensure that the appropriate traffic-monitoring criteria are met. Potential applications for the systems include congestion charge monitoring and other road toll applications, traffic surveys, and monitoring of secure car parks.
There are many methods of locating patterns in images, however, no one approach is best for all applications. Typical pattern-search techniques include
* Normalized gray-scale correlation
* Geometric-based search
* Binary chain-code-based search
* Decision-tree-based search
* Neural-based search algorithms
Firstsight Vision's pattern-recognition tool, Manto, is part of the Common Vision Blox hardware-independent machine-vision toolkit from Stemmer Imaging (Puchheim, Germany; www.imaging.de). Since each pattern-search technique has its advantages and disadvantages, Manto combines the positive characteristics of existing tools, allowing completely new application areas to be addressed.
One result of this approach is that the tool allows the identification of objects with organic fluctuations in their composition, such as different types of fruit, gender differences in human faces, and the identification of defects in surface textures. Most important, however, the versatile recognition characteristics of Manto can be applied to real-time recognition of vehicle types that would not be possible using the individual techniques listed above.
The system uses the following features of an image: correlation, geometrical connections, texture, and color. A nonlinear multiresolution filter is used to independently transform the picture, determine the relevant features of an object, and enter these in a feature vector. A neural network then separates the object classes on the basis of these features. The neural technology used here is known as a Support Vector Machine (SVM).
SVMs work by creating a decision surface with optimal generalization ability (that is, the ability to recognize objects not contained in the training set). Essentially the system learns to identify the patterns of interest from a set of training images and then calculates a confidence factor for its classification choice for each test image. For traditional neural networks, there is a finite number of training images that the system can deal with--a phenomenon known as "over fitting" When this point is reached, the effort to separate classes becomes too difficult and error rate saturation occurs.
As the number of sample images increases, the error rate drops until it reaches a saturation point and does not improve any further than this. The SVM approach does not suffer from over fitting problems, allowing the use of any number of training images, providing sufficient memory can be allocated for the fitting process. As a result, objects can be classified with accuracy not previously possible.
The system has been used in a pilot study on a relatively small number of samples. Images of about 900 vehicles were used for training purposes, and the resulting classifiers used on a similar number of vehicles. In view of the relatively small number of vehicles being tested, the vehicles were also counted and classified manually to compare with the results from Manto. An accuracy rate of 85% was achieved, and it is anticipated that 95% is a realistic target.
The confidence factor for each classification is between 0.5 and 1 (1 being 100% confident). The system can be used with any camera and any control system, that is, it could be used with a 'smart' camera or with a PC-based control system. A variety of cameras can be used for these type of applications, including megapixel digital cameras with auto iris and burst trigger control, which produce high-resolution traffic images even in situations in which the lighting conditions are uncertain.