Optimising Baseline Distance in Stereo Cameras: An Experimental Approach to Enhance Object Distance Accuracy

This research explores how the variation of baseline distance influences the accuracy of object distance estimation in stereo camera systems based on a self-collected stereo image dataset. The dataset contains stereo images of people recorded during a laboratory session where participants were placed at distances between 2 and 50 meters, while the baseline distances could be set between 25 cm to 275 cm. While this dataset offers optimal and stringent evaluation conditions, it suffers from the notable drawback of being restricted to a single object category (human subjects) and relatively low levels of dynamism for the surrounding environment. A quantitative experimental methodology was implemented that included object detection by Mask R-CNN and depth estimation through disparity maps using Stereo Semi-Global Block Matching (SGBM). Results indicated a clear correlation between baseline distance and measurement accuracy, with the 75 cm baseline achieving optimal performance (MAE: 0.23 m; RMSE: 0.34 m). Conversely, a shorter baseline (25 cm) resulted in significant errors, especially for distant objects (MAE: Measurements taken using baselines longer than 175 cm showed increased errors when measuring short distances. The RMSE value at a 75 cm baseline was measured at 0.34 m, which corresponds to only 0.68% of the maximum tested distance (50 m). This result highlights the effectiveness of the experimental design and approach, as it is significantly lower than the commonly recognized RMSE threshold for stereo-based distance estimation. Nevertheless, additional validation using more different data records and more complex environmental conditions is extremely important to improve the reliability and applicability of these results for actual use exceeding the limits of current data records. Keywords: stereo camera, baseline, object distance, mask R-CNN, disparity map