Architectural Adaptation for road humps detection in adverse light scenario

Road hump is a semi-cylindrical elevation on the road made across specific locations of the road. The vehicle needs to manoeuvre the hump by reducing the speed to avoid car damages and pass over the road hump safely. Road Humps on road surfaces if identified in advance helps to maintain the security and stability of vehicles, especially in adverse visibility condition viz. night scenarios. We have proposed a deep learning architecture adaptation by implementing MISH activation function and developing a new classification loss function called “Effective Focal Loss” for Indian road humps detection in adverse light scenario. We captured images comprising of marked and unmarked road humps from two different types of cameras across South India to build heterogeneous dataset. Heterogeneous dataset enabled the algorithm to train and improve the accuracy of detection. The images were pre-processed, annotated for two classes viz, marked hump and unmarked hump. The dataset from these images was used for training the single stage object detection algorithm. We utilised an algorithm to synthetically generate reduced visible road humps scenario. We observed that our proposed framework effectively detected the marked and unmarked hump in the images in clear and adverse light environment. This architectural adaptation sets up an option for early detection of Indian road humps in reduced visibility condition thereby enhancing the autonomous driving technology to handle wider range of real-world scenario.