Efficient Attention Mechanism and Perspective Transform in Realtime Semantic Segmentation

Abstract

Pedestrian guidance systems are essential for assisting individuals with cognitive impairments and enhancing urban mobility. In this paper, we propose an efficient attention mechanism and perspective transform for real-time semantic segmentation to improve the performance of pedestrian guidance systems. We utilized the STDC2 semantic segmentation machine learning model as our base model and integrated the CBAM and to enhance its efficiency and accuracy. Although the addition of these attention modules led to increased accuracy, the model's accuracy-computation tradeoff was not significantly improved. To address this challenge, we investigated the use of the Korean pedestrian dataset for training our model. However, the dataset's viewpoint is downward facing, which is not suitable for a pedestrian guidance system. We employed perspective transformation techniques to adapt the dataset for training a model that can effectively guide pedestrians in urban environments. The proposed combination of an efficient attention mechanism and perspective transform allows for the development of a robust and accurate real-time semantic segmentation model. Our approach aims to enhance the performance of pedestrian guidance systems, providing better assistance to individuals with cognitive impairments and improving overall urban mobility.