In response to the need for real-time detection of surface cracks in automotive steering knuckle fatigue tests, this paper delves into the core module of the crack detection system—the crack recognition and location detection module. This module aims to achieve rapid and precise localization and identification of cracks in collected images. After analyzing mainstream object detection algorithms, YOLOv4 was selected as the foundational framework. Given the specificity of crack targets and the dual requirements of real-time performance and accuracy for detection tasks, two key improvements were proposed: firstly, a data augmentation method based on fractal characteristics of crack morphology, which effectively enhances the model’s generalization ability by simulating the crack propagation process; secondly, optimizing the network structure by introducing various strategies, including decoupling heads, an anchor-free mechanism, a Focus structure, and adjusting network depth and width. Experimental results show that the improved model’s mean average precision (mAP) increased from the baseline of 66.16% to 78.73%, and the detection speed significantly improved from 15.3 FPS to 56.8 FPS, significantly outperforming the original YOLOv4 model. This provides effective algorithmic support for real-time and reliable detection of steering knuckle fatigue cracks.