Please use this identifier to cite or link to this item:
|Title:||Side view car detection with hierarchical COSFIRE filters|
|Abstract:||The problem for this project is to find a suitable configuration for side view car detection and localization. The cars need to be detected in complex scenes, which means that the cars will not always be displayed fully and therefore the full shape of the car is not entirely visible. The method of object recognition which is used is called COSFIRE. Another modification of the COSFIRE filters, called S-COSFIRE, was used as well. Currently, both these methods work based on a single training image. Therefore, the first part of this project involved developing a system whereby both these methods can make use of multiple training images. The default COSFIRE model is a 2 layer system, whereby the Gabor filters’ responses are the input for the COSFIRE filters of the whole prototype. The COSFIRE filters were configured with the shape of a whole car. Meanwhile, a 3 layer system is when the initial Gabor filters’ responses are the input to the first set of S-COSFIRE filters. The input for the last layer of COSFIRE filters is the polar coordinates of the second layer. The fact that there are 3 layers instead of the default 2, forms a hierarchy of S-COSFIRE filters. The keypoints for the S-COSFIRE filters were chosen to be the wheels of the car. The data set which was used is the UIUC side view car data set, which is used as a benchmark by several research papers. During the evaluation part of this project, the number of considered training images was altered so that a fair comparison between both methods could be accomplished. The COSFIRE and S-COSFIRE models were both configured to work with multiple prototypes from the positive training examples. More prototypes were needed for the COSFIRE model, since the shapes of cars varies a lot more than the types of wheels found in the data set. The results showed a significant better performance when using the S-COSFIRE model over the COSFIRE model.|
|Appears in Collections:||Dissertations - FacICT - 2016|
Dissertations - FacICTAI - 2016
Files in This Item:
|1.42 MB||Adobe PDF||View/Open Request a copy|
Items in OAR@UM are protected by copyright, with all rights reserved, unless otherwise indicated.