DepthForge

Effective real-time distance estimation is critical for robotic navigation, especially in dynamic and complex environments. This research presents a novel method for estimating the distance to the nearest object from the camera using a sequence of monocular images. It also features a unique data-collection approach that leverages robotic simulation platforms to rapidly create varied and comprehensive datasets. By analyzing a sequence of five images, our approach classifies the distance to the nearest object into predefined distance classes and demonstrates high accuracy across various simulated indoor environments.

We deploy our custom wheeled-robot platform in the Webots robotic simulation containing realistic indoor environments, for the purpose of data-collection. We place monocular cameras in the front and rear sides of the robot and stack multiple 2D LIDARs along the robot’s vertical axis to capture depth information from various angles. We obtain the distance to the nearest obstacle using these 2d LIDARs and assign classes such as “very near”, “near”, “average”, “far” based on predefined distance ranges.

We randomize the robot’s trajectory by applying differential speeds to the wheels and record a history of five images. The images are captured at regular intervels of 100ms along with their corresponding class labels to obtain the training dataset. To ensure diversity in the training dataset and adaptability to real-world settings, we collect data at different indoor settings such as “Apartment”, “Factory Hall”, “Office” and “Kitchen” with varying levels of lighting and camera FOV.

To evaluate our model, we collect test dataset from a new indoor “Classroom” setting using the same data collection procedure used for training. We report >90% accuracy in the test data, thus validating the efficacy of our novel approach. Further, we collect a minimalistic test dataset from the real world and evaluate our model. While there is a substantial shift in the real world RGB distributions compared to the simulation distribution, our model fares well for “Very Near” and “Near” class labels. We hypothesize that by fine-tuning the model with a minimalistic real-world dataset, our model can adapt to the real-world settings.