Representation Learning Empowered Sensing System for Pedestrian Crowds Perception in Transportation

Abstract: Pedestrians are important yet vulnerable users in modern transportation systems, which account for 50\% of global fatal road traffic injuries. However, most of the existing sensing models, including detection-based, regression-based, and density estimation methods, are developed based on pre-settings and struggle to cope with complicated transportation scenarios. As a result, the paper proposes an ensemble sensing scheme, Scale-Aware Representation Learning Empowered Sensing (SARLES) system, for detecting and sensing pedestrians in various transportation scenarios. The study addresses the common challenges for accurate pedestrian sensing in transportation scenarios, including occlusion, complex backgrounds, scale variation, diverse distribution, perspective changes, small objects, and blurred regions. The system comprises three innovative modules: Encoder-decoder, Density Map Segmentation & Clustering (DMSC), and Local Patch Refinement (LPR). Firstly, the Encoder-decoder module captures global contextual information and handles the scale variability of pedestrians. Secondly, the DMSC module segments the initial density map into multiple local patches based on the density features and generates patches with stable scale and density distributions. Finally, the LPR module utilizes an ensemble FCN network with various kernel sizes to extract accurate local features from every input patch based on its density and scale features for density map refinements. In summary, the integration of three innovative modules addresses the common challenges, making it a powerful tool for detecting and counting diverse distributed pedestrian groups in complex transportation scenarios. Extensive experiments on three challenging benchmarks indicate superior performance in comparison to the existing state-of-the-art methods.