This paper aims to investigate the impact of occasional traffic crashes on the urban traffic network flow. Toward this purpose, an extended model of coupled Nagel–Schreckenberg (NaSch) and Biham–Middleton–Levine (BML) models is presented. This extended model not only improves the initial conditions of the coupled models, but also gives the definition of traffic crashes and their spatial/time distribution. Further, we simulated the impact of the number of traffic crashes, their time distribution, and their spatial distribution on urban network traffic flow. This research contributes to the comprehensive understanding of the operational state of urban network traffic flow after traffic crashes, towards mastering the causes and propagation rules of traffic congestion. This work also a theoretical guidance value for the optimization of urban traffic network flow and the prevention and release of traffic crashes.

To meet the requirement of real-time bridge crack detection, this paper proposes a lightweight detection model based on YOLOv7-tiny. First, an edge-preserved image enhancement method is proposed. It effectively enhances the image contrast and preserves the structural features of crack edges. This provides a high-quality data foundation for the detection network. Second, a LWCSP module is introduced. This module integrates hybrid convolution and shuffle operations. It reduces the model’s parameter count and computation. Simultaneously, it maintains strong feature representation capability. A good balance between detection performance and efficiency is achieved. Finally, an improved SWise-IoU is proposed to optimize the bounding box regression in YOLOv7-tiny. This method dynamically evaluates sample quality. It enables differentiated gradient adjustment for samples of different qualities. This promotes sufficient learning of sample features by the model, thereby improving detection accuracy. Experimental results show that the proposed model delivers strong performance on a public bridge crack dataset. Compared to the baseline, the mAP@0.5 is 12.1 higher, and model size, parameter count, and FLOPs are reduced by 7.3%, 8.03%, and 10%, respectively. The final model size is only 11.4 MB, and mAP@0.5 is 86.1%, suitable for a real-time crack detection task.