Pressures from environmental agencies contribute to the challenges associated with the disposal of waste tyres, particularly in South Africa. Recycling of waste tyres in South Africa is in its infancy resulting in the historically undocumented and uncontrolled existence of waste tyre stockpiles across the country. The remote and distant locations of such stockpiles typically complicate the logistics associated with the collection, transport and storage of waste tyres prior to entering the recycling process. In order to optimize the logistics associated with the collection of waste tyres from stockpiles, useful information about such stockpiles would include estimates of the types of tyres as well as the quantity of specific tyre types found in particular stockpiles. This research proposes the use of computer vision for categorizing individual tyres and estimating the number of visible tyres in tyre stockpile images to support the logistics in tyre recycling efforts. The study begins with a broad review of image processing and computer vision algorithms for categorization and counting objects in images. The bag of visual words (BoVW) model for categorization is tested on two small data sets of tread tyre images using a random sub-sampling holdout method. The categorization results are evaluated using performance metrics for multiclass classifiers, namely the average accuracy, precision, and recall. The results indicated that corner-based local feature detectors combined with speeded up robust features (SURF) descriptors in a BoVW model provide moderately accurate categorization of tyres based on tread images. Two feature extraction methods for extracting features for use in training neural networks (NNs) for tyre count estimations in tyre stockpiles are proposed. The two feature extraction methods are used to describe images in terms of feature vectors that can be used as input for NNs. The first feature extraction method uses the BoVW model with histograms of oriented gradients (HOG) features collected from overlapping sub-images to create a visual vocabulary and describe the images in terms of their visual word occurrence histogram. The second feature extraction method uses the image gradient magnitude, gradient orientation, and edge orientations of edges detected using the Canny edge detector. A concatenated histogram is constructed from individual histograms of gradient orientations and gradient magnitude. The histograms are then used to train NNs using backpropogation to approximate functions from the feature vectors describing the images to scalar count estimations. The accuracy of visible object count predictions are evaluated using NN evaluation techniques to determine the accuracy of predictions and the generalization ability of the fit model. The count estimation experiments using the two feature extraction methods for input to NNs showed that fairly accurate count estimations can be obtained and that the fit model could generalize fairly well to unseen images.
| Identifer | oai:union.ndltd.org:netd.ac.za/oai:union.ndltd.org:nmmu/vital:28313 |
| Date | January 2017 |
| Creators | Eastwood, Grant |
| Publisher | Nelson Mandela Metropolitan University, Faculty of Science |
| Source Sets | South African National ETD Portal |
| Language | English |
| Detected Language | English |
| Type | Thesis, Masters, MSc |
| Format | xvii, 208 leaves, pdf |
| Rights | Nelson Mandela Metropolitan University |
Page generated in 0.0042 seconds