Abstract

Accurately counting the number of dense objects, such as crowds or vehicles, in an image is a challenging and meaningful task widely used in public safety management and traffic flow prediction. The existing CNN-based density map estimation methods are ineffective for extracting the counting features of long-distance queuing vehicles in traffic jams; In addition, these methods do not focus on counting in complex scenes, such as vehicle count-ing in the human-vehicle mixed scenes. To tackle this issue, we propose MSCNet, a novel multi-scale dilated convolution channel-aware deep network for vehicle counting. The proposed network solves the problem of scale variation for long-distance queuing vehi-cles and improves the ability to extract vehicle features in human-vehicle mixed scenes. The MSCNet consists of a front-end module and three functional modules: the front-end module is used to extract the initial features of the counting image; the direction-based per-spective coding module (DPCM) encodes the perspective information of the image from four directions to extract continuous long-distance features; the multi-scale dilated residual module (MDRM) can densely extract the large-scale variation features; the channel-aware attention module (CAM) effectively enhances the channel features that are important for vehicle counting in mixed human-vehicle scenes. The MSCNet has conducted extensive comparative experiments on the TRANCOS dataset, the VisDrone2021 Vehicle&Crowd dataset, and the ShanghaiTech dataset. The experimental results show that the MSCNet outperforms the state-of-the-art counting networks for dense vehicle counting, especially in mixed human-vehicle scenes.

Recommended citation: Qiyan Fu, Weidong Min*, Chubbo Li, Haoyu Zhao, Ye Cao, Meng Zhu. MSCNet: Dense vehicle counting method based on multi-scale dilated convolution channel-aware deep network. GeoInformatica, 2024, 28 (2): 245-269. DOI: 10.1007/s10707-023-00503-7.
Download Paper | Download Slides