Obtaining accurate traffic loads is crucial for the assessment of bridges. The traffic load obtained by the current method is insufficient for the refined analysis of bridge structures. Herein, a fusion method is proposed to generate fine‐grained traffic load spectra using weigh‐in‐motion data, video‐based vehicle spatial–temporal information, and knowledge‐based information of historical passing vehicles. Its effectiveness is tested on an interchange viaduct in Shaanxi, China. The average biases of the longitudinal and transverse locations of driving vehicles, which were identified using the proposed method, are 1.31 and 0.14 m, respectively. The identification accuracy in these two directions improved by 19% and 56%, respectively, compared with that of a pure deep learning‐based video identification method. Meanwhile, the accuracy of identifying the axle number is 99.87%. Additionally, a fine‐grained traffic load spectrum automatically generated with high accuracy is demonstrated. This method can be extended to other scenarios to further analyze and predict vehicle‐related bridge performance.