Real-time Pedestrian Detection and Safety Alert System

Problem

Intersections are most dangerous in the moments rules are broken: a pedestrian stepping into the road against a restricted signal, or a vehicle speeding through a red light. Alerts for those moments only help if they fire immediately, which rules out shipping video to a server — the detection has to run in real time on a small edge device at the intersection itself.

Approach

I built a real-time pedestrian detection system on the NVIDIA Jetson Nano, training YOLO and MobileNet SSD models on the CityPersons and Caltech pedestrian datasets and comparing them for accuracy and speed on edge hardware. On top of the detections, I developed audio-visual alerts that trigger when pedestrians enter restricted roadways and when vehicles speed through red lights.

Architecture & Tech

The models run on the Jetson Nano through NVIDIA's inference stack. To reach real-time frame rates on a device this small, I optimized inference with quantization and pruning, compiled the models with TensorRT, and used DeepStream for the video pipeline. The alert logic sits on the detection stream and drives the audio-visual warnings.

Results & Impact

The system detects pedestrians and triggers its safety alerts in real time entirely on the edge device — no server round-trip — covering both alert scenarios: pedestrians in restricted roadways and vehicles speeding through red lights.

What I'd do next

I would evaluate the system across more intersection geometries and lighting conditions, and measure how much further INT8 calibration and newer lightweight detector architectures could push the frame rate on the same hardware.