Using Covariance Matrices as Feature Descriptors for Vehicle Detection from a Fixed Camera

Boston University

Using Covariance Matrices as Feature Descriptors for Vehicle Detection from a Fixed Camera

Paper submitted in partial fulfillment of the requirements for the Final Project of SC520 : Digital Image Processing and Communication

Kevin Mader and Gil Reese

There are many potential uses for object identification ranging from development of automatic key-wording in a photo library to determining the weight load on a bridge or specific road or possibly estimating air pollution by the different types of vehicles present. The specific goal for this paper is to develop a method to distinguish and count the number of cars and trucks on the road at a given time. The problem of identifying vehicles has been mostly related to tracking uses, but several have proposed interesting approaches to the problem. Methods such as deformable template matching and template differencing (M Betke 1996) have been used for problems that necessitated real-time algorithms. Slower methods using histograms in the wavelet domain were able to detect objects from a variety of viewing angles (H. Schnelderman 2000).

The method we used was derived largely from a paper about covariance matrices as a distance metric (O. Tuzel 2006). However since the specific problem we had of identifying vehicles in a feed of images, we decided to utilize the information that could be obtained for comparing temporally separated frames. So instead of using region growing methods to identify the vehicles we used background subtraction and simple image segmentation methods to identify the regions where vehicles were located. Then we used covariance Copyright c 2006, Kevin Mader and Gil Reese. All rights reserved. matrices on these regions to determine what kind of vehicle if any was present in the bounding box. One of the decisions that must be made before beginning to do any kind of analysis is to determine what are the groupsvehicles are being classified into. Our selection was to define smaller vehicles as cars and larger vehicles as trucks with the cutoff being around the size of a midsize SUV such as a Chevrolet Tahoe. The decision was somewhat arbitrary, but the general reasoning behind it was caused by initially the small number of trucks present in the data set we chose which would be prohibitive of doing extensive testing. Secondly though was the simplicity of the separation of car and semi-truck. The two objects are vastly different in size and demonstrating a complex method using covariance matrices could separate these two classes of vehicles would be minimally informative since a simple pixel counting method could perform the task. Thirdly choosing groups that were closer together could allow us to investigate what kind of shape information in the covariance matrices allows the algorithm to make a decision to classify a vehicle as either a car or truck.

[...]