DMLViT: Dynamic Multi-Scale Local Vision Transformer for Object Counting in Congested Traffic Scenes

Object counting in congested traffic scenes is an important component of traffic perception, facilitating urban traffic management and public transportation capacity optimization. Vision Transformer (ViT) models have achieved remarkable performance in object counting tasks due to their ability to capture long-range dependencies among diverse image blocks. In this paper, we observe that most existing ViT-based models typically employ fixed-size attention, which is insufficient to address the significant scale variation present in traffic scenes. To overcome this issue, we propose an effective Dynamic Multi-scale Local Vision Transformer (DMLViT) which introduces a priori multi-scale representation to effectively handle scale variation. Specifically, we carry out self-attention on multiple windows around each query token and dynamically fused these attention features to model multi-range spatial dependencies. Two dynamic aggregation strategies are designed to assign corresponding weights to different window branches along the channel or spatial dimensions. Furthermore, as the original feed-forward network (FFN) has limitation in model local information, we develop an efficient local FFN to better capture local dense information. Extensive experiments on four crowd counting datasets and three vehicle counting datasets demonstrate that the proposed DMLViT exhibits promising performance compared to state-of-the-art methods.