Efficient and Scalable Highway Asset Detection
Analytics
22 views ◎6 downloads ⇓
Abstract
The roads and highways are a valuable asset for the state department of transportation. It takes massive investment and a huge amount of time to maintain the road assets condition. Therefore, it becomes important to automate the maintenance process with minimum manual inspection. There has been significant research in the domain of classical computer vision techniques and machine learning methods concerning highway and road assets maintenance. However, the time consumed to assess and maintain and the amount of manual inspection involved is considerably large. Although adding automation to speed up the inspection process has been investigated by many studies, they overshadow the importance of scalable and light frameworks detecting the assets like storm drain and drop inlet in real-time. Thus, this thesis focuses on integrating a reliable and scalable AI Deep Learning framework customized for highway assets localization and detection of the assets in a road infrastructure environment mitigating the need for large and bulky model sizes. Furthermore, utilizing the advantage of the less computational cost and the lightweight framework architecture along with reasonably higher accuracy, it is possible to build an end-to-end framework that supports object localization and detection followed by the inference on the mobile edge embedded platforms. In a nutshell, this thesis presents a Deep Learning localization platform, customized to predict the position or the location of the asset items on the Highways such as drop inlets and storm drains. Moreover, it also provides results on the scalability of the localization task to the multi-object detection task with the help the state of the art EfficientDet-D0 model with a test accuracy of 73.4\% mAP on the annotated test dataset and achieving validation accuracy of 51.67\% mAP on the customized merged data of highway asset item drop inlet and 5 other COCO classes for object detection application. Additionally, various analyses based on mIOU and classification scores are described in the experimental section below. It also represents that the model framework is edge deployable friendly and can be quantized to an Fp16 lighter version of the model with help of the NVIDIA TensorRT engine showing the benchmarking performance of 50.55 FPS on the NVIDIA Jetson AGX Xavier mobile embedded platform. Moreover, it highlights the challenges and the future scope expansion of the work to the real-time onboard drone visual analytics for Highway assets defect detection.