PROJECT TITLE :

SWNet A Deep Learning Based Approach for Splashed Water Detection on Road

ABSTRACT:

Unfavorable weather conditions pose a significant risk to the public's safety on the roads, and this is especially true during periods of rainfall, when water pools on the pavement and increases the risk of automobile collisions, as well as injuries to pedestrians and even fatalities. The automatic detection of splashed water based on Surveillance videos is an interesting and potentially useful method for effectively preventing traffic accidents. Surveillance videos, on the other hand, have a great deal of variation, including shifting lighting, varying illumination conditions, and complex backgrounds, all of which make automatic recognition extremely challenging. In this paper, a novel approach to detecting splashed water based on deep learning is proposed. The approach is presented. To the best of our knowledge, this is the first piece of literature on this subject that makes use of deep learning. An efficient Semantic Segmentation network known as SWNet has been innovatively proposed as a means of isolating the potential regions in which splashed water occurred. To record the distinct optical qualities of water that has been splashed, a structure called an encoder-decoder has been developed. SWNet is able to operate at a high level of efficiency because it recycles pooling indices and uses a lightweight decoder. The multi-scale feature fusion structure that SWNet possesses allows for the integration of both the coarse semantic information and the detailed appearance information. This results in a significant increase in accuracy as well as a refinement of the edge segmentation. In order to account for the imbalanced distribution that exists between splashed water and backgrounds, a weighted cross entropy loss has been implemented for the splashed water. In addition, a splashed water attention module has been developed to concentrate on the important parts of moving vehicles and splashed water by means of an Attention Mechanism that integrates global contextual information in semantic segmentation. This is accomplished by paying attention to the salient regions of the image. The effectiveness and efficiency of the proposed method, which outperforms the methods that are considered to be state-of-the-art, was demonstrated through experiments that were conducted on a newly collected splashed water dataset.