| dc.description.abstract | Given the labour-intensive and subjective nature of manual sewer assessments
using closed-circuit television (CCTV), computer vision techniques offer a
practical alternative to human effort in sewer inspections. However, existing
advancements have primarily focused on classifying CCTV frames and
detecting or segmenting defects within those frames. This thesis presents an
integrated framework for automated CCTV-based sewer condition assessment,
consisting of a deep learning (DL)-based method to address defect identification
and combined image processing techniques for subsequent severity
quantification based on the output of identification. The identified defects are
categorized into operational defects and structural ones, with several defect
classes selected as representative examples in experiments, and specific
methods for quantifying their severity are designed separately in this study.
Multiple real CCTV images containing the targeted defects are then used to
demonstrate the performance of the proposed methods.
Regarding defect identification, with knowing the advancements in deep neural
network (DNN) architectures and the recognized significance of data in artificial
intelligence, this research emphasizes a data-centric analysis of DL-based
sewer defect detection, using a DNN-based object detector, YOLOv4, to detect
settled deposits as an example. The experiments demonstrated that increasing
the training dataset size improves the model's performance up to a certain
threshold. The findings also underscored the importance of consistent dataset
annotation. While image augmentation and transfer learning can enhance
average precision, they have varied effects on other evaluation metrics.
A new approach for quantifying the severity of operational defect via estimating
pipe cross-section area loss was developed in this thesis, and settled deposit
was selected as a representative operational defect to be studied. The
proposed assessment approach includes a joint fitting module and a defect
segmentation module, which are then integrated and transformed to eventually
quantify the loss of cross-sectional area. The joint fitting module involves
detecting a valid joint in the footage, applying a combination of image processing techniques, and using an ellipse fitting algorithm. Defect
segmentation is achieved through K-means clustering, complemented by
customizing a series of image processing operations. In the final quantification
of severity, the distance between the defect and the joint was taken into account.
Each module is evaluated on several real-world cases. The evaluation results
show that the joint fitting module can generate ellipses that align with the
outlines of joints in most test cases, with minor deviations mainly due to joint
displacement. Together with the designed defect segmentation module which
identifies the regions of settled deposits, the whole framework results in
estimated cross-sectional area loss values that are consistent with visual
assessments. Additionally, the findings highlight the necessity of converting the
fitted ellipse area to the pipe’s cross-sectional area at the defect location for
accurate cross-sectional area loss calculations.
This study also makes contribution to the automated assessment of structural
defects, which is divided into two branches. The first focuses on joint-related
defects, especially choosing the validation of detected joint displacement as
research target, and customized approach is built via integrating image
processing techniques with morphological analysis. The proposed method is
demonstrated through various real CCTV images, and it successfully estimates
joint displacement distances that align with human estimations. The second
branch concentrates on classifying cracks and fractures. The developed method
stacks various image processing operations to enhance and extract
morphological features, followed by a classification approach that integrates
polar transformation and principal component analysis. It is also tested with
several real cases under various conditions, and the results shows it can
successfully extract the main morphological features of cracks and fractures
and eventually classify them into different groups corresponding to different
condition deduct values.
Lastly, this thesis offers recommendations for future work. Firstly, the
assessment methods could be further enhanced by incorporating more complex
scenarios during their development. At the same time, significant attention
should also be directed towards the defect identification model, as it serves as the foundation of the entire framework. Moreover, additional modules should be
developed for the full automation in practical applications. These may include,
but are not limited to, evaluating varying camera angles, counting the number of
defects, assessing other defect types not covered in this study, and
automatically generating comprehensive assessment reports. | en_GB |
