Objective: This study aimed to develop and validate a fully automated, deep-learning framework for the precise and standardized measurement of lumbar intervertebral range of motion (IROM) and sagittal translation (∆ST) from dynamic radiographs, to overcome the labor-intensity and high inter-observer variability of manual methods. Methods: We conducted a multicenter retrospective study using a dataset of 905 patients (1810 images) from three institutions. A Multi-task High-Resolution Network (HRNet) was developed, utilizing an encoder for multi-scale feature extraction and two decoders for the automated identification of vertebral body centroids (L1-S1) and localization of vertebral corners. IROM and ∆ST were then automatically calculated from these landmarks. Model performance was trained and validated on an internal set of 748 patients and tested on an external set of 157 patients. Results were compared against a gold standard established by expert consensus, with evaluation metrics including identification rate (IR), mean absolute error (MAE) for landmark detection, and MAE for the kinematic parameters. Results: The model demonstrated outstanding performance. On the internal (external) test sets, the centroid identification rate was 98% (98%), and the MAE for corner localization was 2.13 mm (2.37 mm). Clinical parameter measurement showed excellent alignment with radiologists, with MAEs of 2.98° (3.14°) for IROM and 1.67 mm (1.64 mm) for ∆ST on the internal (external) datasets, revealing no significant systematic bias. The automated system drastically improved efficiency, reducing average processing time per patient from 180 seconds to 0.1 seconds. Conclusion: The proposed Multi-task HRNet provides a highly precise, efficient, and objective solution for automated lumbar kinematic analysis. Its robust performance across multicenter data and strong clinical alignment support its potential as a viable tool for standardizing the assessment of spinal mobility and instability in clinical practice.