Classifying Road Debris Using Deep Learning Technique in Artificial Intelligence
Disciplines
Computer-Aided Engineering and Design | Heat Transfer, Combustion | Manufacturing
Abstract (300 words maximum)
According to a study done by AAA Foundation for Traffic Safety in 2016, road debris was a factor in an average number of 50,658 police-reported crashes between the years 2011-2014. This work addresses the critical problem of road debris detection and classification, a major threat to road safety, especially on highways. Road debris, such as barrels, car parts, puddles, salts, and trees, can cause accidents. Leveraging deep learning, we explored three pre-trained convolutional neural network (CNN) models - VGG16, MobileNetV2, and InceptionResNetV2 - to classify five types of road debris. We divided our dataset into training, validation, and testing sets, initially with 146, 73, and 49 images. After augmenting the dataset, we increased it to 875 training thermal images, 375 validation thermal images, and 114 testing thermal images. We evaluated the models' performance over various epochs with a learning rate of 0.0001, an Adam optimizer, and a batch size of 10. The VGG16 model emerged as the top performer, boasting a 100% training accuracy and a 96.65% validation accuracy. In testing, it correctly classified 90.35% of the images. Visualized confusion matrices showed consistent superiority for the VGG16 model across all debris types. This work underscores the efficacy of deep learning models in detecting and classifying road debris, with VGG16 as the most accurate and efficient model. It also emphasizes the importance of image augmentation, significantly improving model performance by expanding the training dataset's size and diversity. These findings have substantial implications for road safety. Implementing deep learning models for road debris detection can substantially reduce accidents, making roads safer for all users. Road authorities and safety organizations can leverage this research to develop automated systems for timely debris detection and removal, enhancing road safety.
Academic department under which the project should be listed
SPCEET - Mechanical Engineering
Primary Investigator (PI) Name
Sathish Gurupatham
Classifying Road Debris Using Deep Learning Technique in Artificial Intelligence
According to a study done by AAA Foundation for Traffic Safety in 2016, road debris was a factor in an average number of 50,658 police-reported crashes between the years 2011-2014. This work addresses the critical problem of road debris detection and classification, a major threat to road safety, especially on highways. Road debris, such as barrels, car parts, puddles, salts, and trees, can cause accidents. Leveraging deep learning, we explored three pre-trained convolutional neural network (CNN) models - VGG16, MobileNetV2, and InceptionResNetV2 - to classify five types of road debris. We divided our dataset into training, validation, and testing sets, initially with 146, 73, and 49 images. After augmenting the dataset, we increased it to 875 training thermal images, 375 validation thermal images, and 114 testing thermal images. We evaluated the models' performance over various epochs with a learning rate of 0.0001, an Adam optimizer, and a batch size of 10. The VGG16 model emerged as the top performer, boasting a 100% training accuracy and a 96.65% validation accuracy. In testing, it correctly classified 90.35% of the images. Visualized confusion matrices showed consistent superiority for the VGG16 model across all debris types. This work underscores the efficacy of deep learning models in detecting and classifying road debris, with VGG16 as the most accurate and efficient model. It also emphasizes the importance of image augmentation, significantly improving model performance by expanding the training dataset's size and diversity. These findings have substantial implications for road safety. Implementing deep learning models for road debris detection can substantially reduce accidents, making roads safer for all users. Road authorities and safety organizations can leverage this research to develop automated systems for timely debris detection and removal, enhancing road safety.