MIT Senseable City Lab

Favela 4D: Deep Learning for Slum Expansion Analysis

Utilized deep learning approaches to analyze satellite imagery data, monitoring the three-dimensional expansion of favelas in Rio de Janeiro.

Spearheaded outward expansion analysis by benchmarking semantic segmentation models (Unet/ResUnet/Segformer) on satellite imagery, achieving 79% IoU accuracy (benchmark: ~75%). Testing deep learning, unsupervised, and generative AI methods (ongoing).

Developed an LSTM-based time series prediction model using 20-year favela boundary data (1999-2019) to predict expansion trends, enabling data-driven urban planning decisions.

Collaborated with Lidar members to develop 3D expansion segmentation through satellite-point cloud fusion.[ry1] 

Metabolic Walking: Computer Vision for Pedestrian Behavior Analysis Under Varying Weather Conditions

Trained a shadow detection segmentation model with 95% accuracy, enabling precise classification of shaded areas in urban environments.

Developed an automated pedestrian tracking model to extract movement patterns (walking speed, presence in shaded areas, demographic attributes), providing quantitative insights on pedestrian behavior in relation to shade.

Analyzed pedestrian response to shading infrastructure under varying weather conditions, supporting adaptive urban design and mobility policy.

work in progress .. limited illustration due to confidential concern