A smart transportation system aims to steer vehicles without encountering accidents on the way to the destination. Autonomous vehicles greatly rely on the camera to perceive their dynamically changing surroundings. Deep learning solutions to vision data are driving the research on collision mitigation systems for self-driving cars or automobiles with a higher level of autonomy. This paper presents a three-class vehicular collision image classification dataset created from crowd-sourced dashboard camera videos containing on-road vehicular collisions. An image classification deep learning model is developed using “Teachable Machine” by transfer learning on the proposed dataset that consists of 8729 images across three classes, namely “No Collision”, “Collision”, and “Collided”. The ability of the developed model to classify vehicular collision images based on the spatial aspect of the crash is evaluated and presented. The developed model produced a per-class accuracy of 73.13, 73.13, and 78.75% in classifying images and area under the receiver operating characteristics (ROC-AUC) curve of 0.74, 0.68, and 0.73 is obtained for images containing no collision, the occurrence of a collision, and after the occurrence of collision, respectively.
On-Road Vehicular Collision Image Classification Using Deep Learning
Lect. Notes in Networks, Syst.
International Conference on Computing and Machine Learning ; 2024 ; Rangpo, India March 29, 2024 - March 30, 2024
2024-10-23
13 pages
Article/Chapter (Book)
Electronic Resource
English
Vehicular Collision Avoidance at Intersection Using V2I Communications for Road Safety
Springer Verlag | 2020
|VEHICULAR COLLISION SENSOR AND VEHICULAR COLLISION DETECTING DEVICE USING THE SAME
European Patent Office | 2017
|Modeling vehicle collision instincts over road midblock using deep learning
Taylor & Francis Verlag | 2023
|VEHICULAR COLLISION AVOIDANCE SYSTEM AND VEHICULAR COLLISION AVOIDANCE METHOD
European Patent Office | 2018
|