COSMOS Smart Streetscapes

Smart City Intersections

The persistent COSMOS deployment at Amsterdam Ave and 120th St—equipped with street-level/bird's-eye cameras, wireless connectivity, and programmable edge/cloud resources—provides a unique testbed for city-scale intelligent systems. Vehicles, cyclists, and pedestrians interact under occlusion, changing lighting and weather, heterogeneous sensing conditions, and safety-critical constraints. The COSMOS smart-intersection research program combines programmable multi-view cameras, wireless connectivity, edge/cloud computing, and long-term real-world deployments to study the complete pipeline from urban sensing to intelligent decision making.

Our research has evolved from real-time object detection and tracking toward data-driven traffic modeling, digital twins and controllable simulation, semantic traffic understanding, and adaptive urban AI. Long-term observations from the COSMOS intersection support the development of robust perception models, realistic traffic simulations, and synthetic safety-critical scenarios. More recent work investigates vision-language and language-based reasoning for understanding complex traffic interactions, as well as adaptive methods for allocating sensing, communication, computation, and annotation resources across heterogeneous data sources.

Together, these capabilities make the COSMOS intersection a testbed not only for sensing and communications, but also for studying how AI systems can observe, model, generate, understand, and adapt to complex urban environments.

Recent work investigates vision-language and language-based reasoning for understanding complex traffic interactions, leveraging zero-shot image tagging (Ho et al., 2026) to transform semantic descriptions into structured traffic evidence.

Real-Time Perception and Edge Analytics

Early research at the COSMOS smart intersection established the sensing, computing, and perception infrastructure for intelligent urban applications (Raychaudhuri et al., 2020; Yang et al., 2020). Using street-level and bird's-eye cameras connected to COSMOS edge and cloud resources, we developed real-time pipelines for detecting and tracking vehicles, pedestrians, and cyclists at Amsterdam Avenue and West 120th Street (Turkcan et al., 2024; Ghasemi et al., 2023; Ghasemi et al., 2022).

Recent advances expand this perception foundation with calibration-free, view-agnostic 3D object detection across ground-level, infrastructure, and aerial viewpoints (Turkcan et al., 2026), alongside real-time video analytics optimized for deployment across edge nodes and end-user devices (Ghasemi et al., 2025). We distinguish between methods evaluated on recorded intersection data and demonstrated live deployments running on testbed hardware.

Together, this work established the core sensing → communication → edge/cloud computing → perception pipeline of the COSMOS smart intersection. Persistent perception produces trajectories, which support spatial/temporal distributions and learned traffic models, enabling recent research to model and generate traffic behavior, understand complex interactions, and adapt data collection and computing to support robust urban AI.

Digital Twins and Controllable Traffic Generation

Long-term observations from the COSMOS intersection enable us to build data-driven models of urban traffic and reproduce complex interactions in controlled environments. Rather than relying only on manually specified traffic rules, our work learns spatial, temporal, and trajectory distributions directly from real-world observations. Persistent perception produces trajectories, which support spatial/temporal distributions and learned traffic models, providing a foundation for realistic traffic simulation and digital-twin experimentation.

Our digital twin research is structured around three core pillars:

1. Data-Driven Trajectory Simulation: Recorded trajectories are modeled as spatial and temporal distributions, sampled to generate new traffic scenes, and refined using learned trajectory-prediction models (Zang et al., 2024).

2. Data-Driven Trajectory Simulation: Recorded trajectories are modeled as spatial and temporal distributions, sampled to generate new traffic scenes, and refined using learned trajectory-prediction models (Zang et al., 2024).

3. Language-Guided Interaction Generation: High-level natural language descriptions of rare and safety-critical events are translated into behavioral segments and physically plausible trajectories (Zang et al., 2026).

Together, these efforts connect the physical COSMOS deployment with increasingly flexible digital representations, enabling researchers to replay observed traffic, synthesize new conditions, and systematically test intelligent transportation systems under controlled and safety-critical scenarios.

Adaptive Data Curation, Communication, and Edge Intelligence

A large-scale urban sensing system produces far more data than can be continuously transmitted, processed, and annotated. Different cameras observe different viewpoints, traffic patterns, occlusions, and environmental conditions, making uniform resource allocation inefficient. To address this, our research is organized around a central decision question: Where should we spend limited sensing, bandwidth, compute, and annotation resources?

We group our edge intelligence initiatives by key system decision variables:

1. What to collect & label: Under a limited annotation budget, adaptive data collection algorithms continuously evaluate model performance and direct labeling resources toward camera views and conditions where new data is most valuable (Zang et al., 2025).

2. What to transmit & where to process: Adaptive video streaming and edge filtering reduce bandwidth and processing costs while preserving information for downstream tasks (Ghasemi et al., 2025).

3. How to manage resources: Distributed vision-language inference partitions model processing between edge devices and cloud servers to optimize execution costs and latency (Li et al., 2025).

More broadly, this research treats data acquisition, communication, computation, and model training as interconnected components of the same system. The goal is to move from a passive sensing infrastructure toward an adaptive urban AI platform that determines what information should be collected, where it should be processed, and how limited resources should be used to improve system-wide performance.