Derek C. Schuurman

A comparison of PCA and ICA for object recognition under varying illumination

An experiment is performed to evaluate the ability of two different subspace methods to recognize objects under different illumination conditions. The principal component analysis (PCA) and independent component analysis (ICA) are compared for classifying 25 different objects with varying degrees of specularity under different illumination. Each object was sampled under three widely different lighting conditions to form a set of training images used to create subspaces with dimensions ranging from 10 to 30 basis vectors. The efficacy of ICA and PCA to correctly classify the objects was tested using two test images for each object under unique lighting conditions not included in the training set. The results were also determined when the images were pre-filtered with a Laplacian of Gaussian filter. Results show that ICA techniques show promise for object recognition under varying illumination conditions.

Parallel architecture for PCA image feature detection using FPGA

This paper describes a parallel architecture for image feature detection implemented using an FPGA. The image features are detected using a localized PCA (principle component analysis) pattern matching scheme. An offline training phase identifies sub-windows surrounding salient points in an object which are then projected into eigenspace. Sub-windows from an input image can then be projected into the same eigenspace in order to recognize the same feature points in other images. An FPGA is developed to sequentially project a 10times10 sub-window surrounding each and every pixel into eigenspace so that features can be detected in an image. The FPGA uses parallel dot-product blocks with parallel multipliers and parallel comparators to enable rapid feature detection for sub-windows. Simulations are performed to determine the feasibility of using an FPGA along with the number of required logic elements and the timing requirements.

Towards real-time sorting of recyclable goods using support vector machines

The amount of recyclable material being processed worldwide is increasing. There is a demand for new technologies that can quickly sort these materials for maximum purity while maintaining high throughput. This paper proposes a method to automatically sort two materials: polycoat containers and PET (Polyethylene Terephthalate) bottles. This method utilizes a visible light camera and does not rely on Near-Infrared spectrometry. This paper proposes a high-speed method to automatically locate regions that likely contain these materials within the image and removes them from the background. These regions are merged into whole containers and are classified as either a polycoat container or PET bottle. This is accomplished using a linear support vector machine (SVM) trained on the histogram of pixel intensities. The proposed method obtained a 93% recognition rate, and is able to run at high frames rates in real-time using a field-programmable gate array.

Step-by-step design and simulation of a simple CPU architecture

This paper describes a sequence of assignments, each building upon the next, leading students to a working simulation of a simple 8-bit CPU (Central Processing Unit). The design features a classic Von Neumann architecture comprising a simple data path with a few registers, a simple ALU (Arithmetic Logic Unit), and a microprogram to direct all the control signals. The first step involves the design of the ALU which is capable of eight basic operations. The second step guides students to construct a datapath complete with several 8-bit registers. The third step involves the design and implementation of a control unit which uses a microprogram to implement machine code instructions. The microprogram implements nine basic machine language instructions which are sufficient for writing many simple programs. The final step involves adding program memory and an input and output device to form a simple working simulation of a computer. At this point, students may hand-assemble code for their CPU and simulate its execution. All simulations are performed using a free and open source simulator called Logisim which performs digital logic simulations with the ability to build larger circuits from smaller subcircuits. Students can set an adjustable clock rate and observe the internal CPU state and registers as it retrieves instructions and steps through the microcode. The basic CPU architecture provides many opportunities for more advanced exercises, such as adding an instruction fetch unit, adding pipelining, or adding more machine language instructions. The assignments were introduced in a second year course on computer organization, providing an effective hands-on approach to understanding how a CPU actually operates.

Vision-based obstacle detection using a support vector machine

This paper describes a monocular vision-based obstacle detection method for a mobile robot using a support vector machine (SVM). A single camera is mounted on the front of a mobile robot and an SVM is trained to classify obstacles as they are encountered by the robot. Since it is not possible to train on all obstacle types a-priori, a one-class SVM is used to learn the appearance of the floor in the absence of obstacles. Anything that is not recognized as a floor is classified as an obstacle. To improve robustness in recognizing floor features, images are preprocessed using a Fast Fourier Transform (FFT) to provide translation invariance. Experimental results indicate high accuracy and specificity for four different floor surfaces that were tested.

Visual sorting of recyclable goods using a support vector machine

Mounting environmental concerns and changing attitudes have led to recycling programs to divert waste from entering landfill sites. This trend has led municipalities to explore improved methods and tools such as machine vision for sorting and managing the growing volume of recyclable materials. This paper describes an approach to visual sorting using image intensity data and a support vector machine applied to the unique problem of sorting polycoat containers from plastic bottles. The approach is rotation, translation and scale invariant since it uses features derived from image histograms. We also demonstrate that the approach is robust to the size, shape, varied labeling and deformation of the recycled material. An experiment is performed to verify the approach using separate test and training data. Despite the use of a modest number of training images, the system achieves a classification accuracy of over 96% using images obtained from a single grey-scale camera.

Video-rate eigenspace methods for position tracking and remote monitoring

The use of principal component analysis is employed for visual position determination and simultaneously for remote visual monitoring. The position of a simple planar robot is visually tracked at video rates using eigenspace methods. The eigenspace image coefficients are simultaneously sent over the Internet to visually display the robot operation at a remote location. A set of basis eigenvectors are first determined using the Karhunen-Loeve Transform (KLT) using an off-line learning process. Once the learning phase is complete, the run-time performance of the eigenspace methods are shown to be fast enough to operate at video rates using off-the-shelf components. The eigenspace provides a compact representation that can be employed for rapid position determination and to provide minimum image reconstruction error for a given number of basis vectors. The computational speed, accuracy, and latency for position determination are experimentally determined. The experimental results show that the eigenspace methods perform well for position tracking and for remote monitoring.

Direct visual servoing using network-synchronized cameras and Kalman filter

A distributed camera system using off-the-shelf components is presented that demonstrates the capability to perform high-speed vision feedback suitable for applications such as direct visual servoing. The limitation of 60 Hz video sample rates is overcome by using multiple RS-170 cameras synchronized over a network to capture at different instants in time. Each camera node has its own computer that processes video at field rates to determine the pose of a planar robot joint using eigenspace methods. Position information is fed back over a network to a master computer to perform direct visual servoing. The resulting vision feedback from the multiple cameras uses a Kalman filter to estimate position and to model the vision computation and transport delays. Computer simulation results are provided as the number of cameras are varied. Finally, real-time experimental results are presented that verify the approach using a network of four cameras performing direct visual servoing of a simple planar robot.

Robust Subspace Position Measurement Using Localized Sub-Windows

The use of localized principal component analysis is examined for visual position determination in the presence of varying degrees of occlusions. Occlusions lead to substantial position measurement errors when projecting images into eigenspace. One way to improve robustness to occlusions is to select small sub-windows so that if some sub-windows are occluded, others can still accurately identify position. The location of candidate sub-windows are predetermined from a set of training images by subtracting the average image from each and then selecting regions using an attention operator. Since attention operators can be computationally time-intensive, the location of all sub-windows are determined a-priori during the training phase. The sub-windows in each of the training images are then projected into eigenspace. Once the training phase is complete, the run-time execution can be performed efficiently since all the sub-windows have been preselected. Input images are classified by each sub-window; majority voting is then used to determine the position estimate. Various experiments are performed including linear and rotational motion, and the ego motion of a mobile robot. This technique is shown to provide greater position measurement accuracy in the presence of severe occlusions as compared to the projection of entire images.