<p>Bolts are widely applied in critical sectors such as industrial manufacturing, construction, and transportation, where they play a key role in large-scale mechanical equipment. The fastening condition of bolts directly affects the stability and safety of equipment operation. To address the challenge of timely detecting bolt loosening, this paper proposes a real-time bolt loosening detection method based on an improved YOLOv8 algorithm combined with machine vision technology. Through comparisons with other object detection algorithms, including Faster R-CNN, EfficientDet, SSD, and DETR, YOLOv8 was selected as the baseline model due to its superior performance in balancing detection accuracy and computational efficiency. The YOLOv8 model was further optimized through lightweight redesign and pruning, reducing the parameter count to 13% and GFLOPs to 31% of the original, while maintaining 95% F1 score and 97% mAP50. The optimized model achieves a processing speed 1.6 times that of the original YOLOv8, with ONNX-format deployment reaching 2.9 times faster inference. After bolt detection, the SIFT algorithm is used to extract image feature points, which are then tracked using the Lucas–Kanade optical flow method. Combined with RANSAC and a distance threshold, mismatched points are filtered out to accurately calculate the bolt loosening angle. To address tracking failures caused by illumination variation and fast motion, an adaptive decision mechanism is introduced, dynamically triggering re-detection and ROI updates based on the number of tracked feature points and the estimated rotation angle. The proposed method is deployed on a Raspberry Pi platform, achieving real-time bolt rotation angle estimation. Experimental results demonstrate the algorithm's robustness under varying lighting conditions, with tracking accuracy exceeding 95% within appropriate parameter settings.</p>