As a key safety component in the automotive chassis system, the fatigue performance of the steering knuckle under complex alternating loads is directly related to the overall vehicle safety. Traditional crack detection methods rely on manual post-observation, which have problems such as low efficiency, easy missed detection, and inability to record the crack initiation and propagation process in real time, making it difficult to meet the needs of modern automotive industry for in-depth research on the fatigue damage law of components. This article systematically analyzes the core requirements for real-time detection of cracks in automotive steering knuckle fatigue tests, and designs an intelligent detection scheme that integrates deep learning, dynamic tracking, and statistical learning. The scheme decomposes the detection task into three core modules: “crack identification and localization”, “dynamic tracking and association”, and “length real-time calculation”. By adopting an improved YOLOv4 object detection algorithm, a dynamic numbering and inheritance mechanism based on similarity criteria, and a crack length estimation algorithm based on regression analysis, real-time, automatic, and accurate detection and measurement of cracks on the surface of the steering knuckle are achieved, providing an effective technical means for studying the fatigue failure law of the steering knuckle.