IOT-ENABLED SMART ADAPTIVE STREET LIGHTING

• 

  System Dashboard

An edge-native, AI-driven street lighting network that learns from traffic and weather to react in under one second while saving 85–90% energy through decentralized, privacy‑preserving intelligence.

Inspiration

Urban lighting is still bound by cloud latency, outages, and one‑size‑fits‑all rules that can jeopardize safety and waste energy; the goal was a system that thinks locally, collaborates globally, and never goes dark.

How we built it

• A hybrid edge‑cloud architecture where each pole hosts an ARM‑based edge unit running a Multimodal Temporal Fusion Transformer for on‑device inference.
• A sensor fusion stack combining cameras, traffic detectors, and weather stations to anticipate demand rather than react to it.
• Federated learning for privacy‑preserving model updates across nodes, meshed over LoRaWAN/Zigbee for self‑healing resilience.
• A lightweight blockchain layer for device identity, auditability, and tamper‑resistant coordination.

What it does

Each lighting node embeds traffic, weather, temporal, and spatial signals, then applies multi‑head attention to produce predictive brightness setpoints with <50 ms inference, enabling <1 s end‑to‑end reactions. Confidence‑scored decision fusion selects safe fallbacks under uncertainty, while federated rounds periodically improve models without exporting raw sensor data. The mesh network maintains service through 30–40% node failures, and the cloud serves only for analytics and backup coordination.

Challenges we ran into

• Edge AI shifts critical latency from seconds to sub‑second, transforming safety behavior at crosswalks and during emergency events.
• Transformer attention heads offer transparent, tunable importance across traffic and weather modalities compared with opaque end‑to‑end CV baselines.
• Federated learning meaningfully cuts bandwidth and privacy risk while keeping accuracy within a few points of centralized training.

Accomplishments that we're proud of

• End‑to‑end response under 1 s; model inference under 50 ms on ARM Cortex‑A72 with NPU acceleration.
• Energy efficiency of 85–90% versus always‑on lighting while preserving illumination quality.
• Target uptime of 99.9% via decentralized control and self‑healing mesh topology.

What's next for IOT-ENABLED SMART ADAPTIVE STREET LIGHTING

• Balancing model size (45–60 MB) with on‑pole compute and thermal constraints required pruning and quantization‑aware training.
• Ensuring privacy with camera inputs demanded on‑device redaction and prohibition of raw frame egress.
• Coordinating federated rounds over lossy links required adaptive scheduling and partial aggregation tolerances.

What we learned

• Edge tier: local transformer inference, actuator control, emergency overrides, and secure key storage.[1]
• Mesh tier: resilient routing, neighbor state sharing, and distributed consensus for group dimming.[1]
• Cloud tier: global analytics, firmware rollout, and last‑resort orchestration without creating a single point of failure.

Why it matters

Cities gain safer, greener streets that continue operating through outages, protect citizen privacy, and improve autonomously—one pole at a time.

Built With

• air?quality;-pwm/0?10v.?-security:-blockchain
• air?quality;-pwm/0?10v.[1]-security:-blockchain
• c/c++
• ci
• containers.?-iot/networking:-zigbee
• containers.[1]-iot/networking:-zigbee
• federated-learning.?-edge:-arm-a72-sbcs-(npu)
• federated-learning.[1]-edge:-arm-a72-sbcs-(npu)
• hw-keys.?-mlops:-qat-+-pruning
• hw-keys.[1]-mlops:-qat-+-pruning
• languages:-python
• linux
• lorawan
• mqtt
• onnx
• onnx/tensorrt
• opencv
• rest/graph-apis
• schedulers.?-data:-edge-emmc;-cloud-time?series-store.?-sensors/actuators:-cameras
• schedulers.[1]-data:-edge-emmc;-cloud-time?series-store.[1]-sensors/actuators:-cameras
• scikit?learn
• secure-boot
• tls.?-cloud:-analytics-+-firmware-rollout
• tls.[1]-cloud:-analytics-+-firmware-rollout
• traffic-+-weather
• typescript.?-ml/frameworks:-pytorch-+-tensorrt
• typescript.[1]-ml/frameworks:-pytorch-+-tensorrt