Background: We study dual-task performance with marker-free video recordings of Timed Up-and-Go tests (TUG) and TUG combined with a cognitive/verbal task (TUG dual-task, TUGdt). Research question: Can gait parameters be accurately estimated from video-recorded TUG tests by a new semiautomatic method aided by a technique for human 2D pose estimation based on deep learning? Methods: Thirty persons aged 60-85 years participated in the study, conducted in a laboratory environment. Data were collected by two synchronous video-cameras and a marker-based optoelectronic motion capture system as gold standard, to evaluate the gait parameters step length (SL), step width (SW), step duration (SD), single-stance duration (SSD) and double-stance duration (DSD). For reliability evaluations, data processing aided by a deep neural network model, involved three raters who conducted three repetitions of identifying anatomical keypoints in recordings of one randomly selected step from each of the participants. Validity was analysed using 95 % confidence intervals (CI) and p-values for method differences and Bland-Altman plots with limits of agreement. Inter- and intra-rater reliability were calculated as intraclass correlation coefficients (ICC) and standard errors of measurement. Smallest detectable change was calculated for inter-rater reliability. Results: Mean ddifferences between video and the motion capture system data for SW, DSD, and SSD were significant (p < 0.001). However, mean differences for all parameters were small (-6.4%-13.0% of motion capture system) indicating good validity. Concerning reliability, almost all 95 % CI of the ICC estimates exceeded 0.90, indicating excellent reliability. Only inter-rater reliability for SW (95 % CI = 0.892;0.973) and one rater's intrarater reliability for SSD (95 % CI = 0.793;0.951) were lower, but still showed good to excellent reliability. Significance: The presented method for extraction of gait parameters from video appears suitable for valid and reliable quantification of gait. This opens up for analyses that may contribute to the knowledge of cognitivemotor interference in dual-task testing.