Selective Scale Context Awareness Network for Object Counting in Intelligent Transportation System

In intelligent transportation systems, object counting is a crucial component of traffic scene perception. The traffic scene always has the properties of strong perspective effects and uneven density distribution, making most existing methods lose their efficiency. To overcome the limitations of existing methods, we develop a selective scale context awareness network (SSCANet) for traffic object counting. Specifically, the proposed SSCANet is combined with a selective separable large-kernel transformer (SSLformer) to deal with scale variation. SSLformer extracts spatial contextual information with selective scale to further establish adaptive long-range dependencies to localize objects with various sizes. To deal with uneven density distribution, the context guidance aggregation module is applied for using global low-resolution contextual information as guidance for high-resolution feature and achieving better feature aggregation. Compared to existing methods, SSCANet demonstrates exceptional performance across three crowd-counting data sets (ShanghaiTech PartA, PartB, and UCF-QNRF) as well as four vehicle-counting data sets (COTRS, CARPK, PUCPR+, and TRANCOS), validating its broad applicability and effectiveness across traffic object counting tasks. Notably, SSCANet exhibits outstanding capability in addressing challenges related to scale variations and uneven density distributions, outperforming mainstream approaches. The introduction of the proposed modules significantly advances the performance of traffic object counting tasks and provides new directions and methodologies for research in this field.