Deterioration of road and pavement surface conditions is an issue which directly affects the majority of the world today. The complex structure and textural similarities of surface cracks, as well as noise and image illumination variation makes automated detection a challenging task. In this paper, we propose a deep fully convolutional neural network to perform pixel-wise classification of surface cracks on road and pavement images. The network consists of an encoder layer which reduces the input image to a bank of lower level feature maps. This is followed by a corresponding decoder layer which maps the encoded features back to the resolution of the input data using the indices of the encoder pooling layers to perform efficient up-sampling. The network is finished with a classification layer to label individual pixels. Training time is minimal due to the small amount of training/validation data (80 training images and 20 validation images). This is important due to the lack of applicable public data available. Despite this lack of data, we are able to perform image segmentation (pixel-level classification) on a number of publicly available road crack datasets. The network was tested extensively and the results obtained indicate performance in direct competition with that of the current stateof-the-art methods.

Condition and deterioration of public and private infrastructure is an issue that directly affects the majority of the world population. In this paper we propose the application of a Residual Neural Network to automatically detect road and pavement surface cracks. The high amount of variance in the texture of the surface and variation in illumination levels makes the task of automatically detecting defects within public and private infrastructure a difficult task. The system developed utilises a feature pyramid core with an underlying feed-forward ResNet architecture. The output from the feature pyramid then feeds into two sub-networks. One sub-network associates a class with the output from the feature pyramid. The other sub-network regresses the offset from each of the output bounding boxes of the feature pyramid to the corresponding ground truth boxes during training. The network was trained on real world data from an already established dataset. The data used to train and test on is very limited, due to the lack of available road crack datasets in the public domain. Despite the limited amount of data, the proposed method achieves a very positive results with minimal error.

This paper details a complete hardware and software system designed to aid in the visual inspection and structural condition monitoring of railway tunnels. The system consists of two main components; an image acquisition system for data collection and an image processing software package for data analysis. The image acquisition system consists of an array of cameras with overlapping fields of view, a uniform lighting solution and a computer to control recording mounted on a rail trolley. The software package carries out operations such as image registration and stitching, 3D reconstruction and change detection. This software package is designed to generate and present image information in a manner which is most useful to structural examiners. The overall system aims to reduce the time taken to carry out visual inspection of tunnels while increasing the overall accuracy of inspections. Our preliminary findings indicate that the system directly benefits examiners when carrying out visual tunnel inspections.

Here we present our multi-core architectures for object detection. We move away from the traditional architecture of Multi-Processors (MPs) by using cacheable accesses to main memory to create atomic cores and utilising local memory for all program data. Main memory is partitioned through software into dedicated data regions to allow atomic accesses by cores, without the need for synchronisation primitives. In doing this, we demonstrate how multi-threading techniques such as Interleaved Task Reordering (ITR) can be utilised to balance the processing loads on available cores. We implement and test up to 7 soft-cores with the Viola Jones face detection algorithm and achieve a performance increase of up to 9.14x with a 100% detection rate: surpassing the theoretical performance increase of multi-core processors for all designs and test images. Furthermore, we surpass the performance increases of multi-core implementations from the literature, thus proving our custom designs to be a more viable solution for multi-core object detection applications. Finally, resource and power consumption estimates indicate our designs to be suitable for embedded systems deployment.

The focus of this paper is a novel object tracking algorithm which combines an incrementally updated subspace-based appearance model, reconstruction error likelihood function and a two stage selective sampling importance resampling particle filter with motion estimation through autoregressive filtering techniques. The primary contribution of this paper is the use of multiple bags of subspaces with which we aim to tackle the issue of appearance model update. The use of a multibag approach allows our algorithm to revert to a previously successful appearance model in the event that the primary model fails. The aim of this is to eliminate tracker drift by undoing updates to the model that lead to error accumulation and to redetect targets after periods of occlusion by removing the subspace updates carried out during the period of occlusion. We compare our algorithm with several state-of-the-art methods and test on a range of challenging, publicly available image sequences. Our findings indicate a significant robustness to drift and occlusion as a result of our multibag approach and results show that our algorithm competes well with current state-of-the-art algorithms.