In response to the technical bottlenecks of poor adaptability to multi-material surfaces and insufficient robustness in complex environments during building crack detection, this study innovatively proposes an adaptive detection algorithm that integrates multimodal features. By reconstructing the YOLOv8 model architecture and combining texture features from visible light images with edge information from frequency domain wavelet analysis, a dynamic feature-weighted pyramid network is constructed to achieve pixel-level localization of cracks on building surfaces, with an average accuracy of 95.4%. An environment-aware enhancement module is developed, which uses a material classification branch based on ResNet-18 and a lighting intensity estimation network to dynamically select either CLAHE enhancement or Retinex correction strategies. This improves the F1-score by 21.8% in strong reflection (>1000 lx) and low contrast (<0.3) scenarios. The system is designed with a pure software architecture, supporting both single-image and video inputs, achieving a mAP@0.5 of 90.7% on the SDNET2018 dataset, providing a lightweight solution for building safety assessment without hardware dependencies.