Predicting Maps Using In-Vehicle Cameras for Data-Driven Intelligent Transport

Abstract

Bird’s eye view (BEV) semantic maps have evolved into a crucial element of urban intelligent traffic management and monitoring, offering invaluable visual and significant data representations for informed intelligent city decision making. Nevertheless, current methodologies continue underutilizing the temporal information embedded within dynamic frames throughout the BEV feature transformation process. This limitation results in decreased accuracy when mapping high-speed moving objects, particularly in capturing their shape and dynamic trajectory. A framework is proposed for cross-view semantic segmentation to address this challenge, leveraging simulated environments as a starting point before applying it to real-life urban imaginative transportation scenarios. The view converter module is thoughtfully designed to collate information from multiple initial view observations captured from various angles and modes. This module outputs a top-down view semantic graph characterized by its object space layout to preserve beneficial temporal information in BEV transformation. The NuScenes dataset is used to evaluate model effectiveness. A novel application is also devised that harnesses transformer networks to map images and video sequences into top-down or comprehensive bird’s-eye views. By combining physics-based and constraint-based formulations and conducting ablation studies, the approach has been substantiated, highlighting the significance of context above and below a given point in generating these maps. This innovative method has been thoroughly validated on the NuScenes dataset. Notably, it has yielded state-of-the-art instantaneous mapping results, with particular benefits observed for smaller dynamic category displays. The experimental findings include comparing axial attention with the state-of-the-art (SOTA) model, demonstrating the performance enhancement associated with temporal awareness.