Inspiration
Road cameras and dashcams generate large amounts of visual data, but most footage remains passive: it is reviewed only after an incident has already happened. MtaaGuard AI is designed to turn existing road video into useful, timely safety intelligence.
What it does
MtaaGuard AI analyzes road video to detect and track vehicles, pedestrians, stalled vehicles, unsafe interactions, road hazards, and near-miss events. It converts visual evidence into structured incident records and severity assessments that can support transport operators, road authorities, schools, gated communities, fleet managers, and safety researchers.
The system follows an agentic perception-decision-action loop. OpenCV 5 processes the video and extracts visual evidence such as trajectories, regions of interest, motion patterns, and event candidates. An orchestration layer then evaluates confidence and severity. When evidence is uncertain, the system can request additional frames or re-run targeted analysis before deciding whether to record the event, recommend an action, or escalate it for human review.
How we are building it
OpenCV 5 is the core vision layer for video decoding, preprocessing, stabilization, perspective transforms, motion analysis, tracking, trajectory analysis, geometric measurements, annotation, and temporal event processing.
AWS provides the cloud execution and delivery layer. Video and generated evidence can be stored in Amazon S3, processed in containerized services running on AWS Graviton, surfaced through an API and dashboard, and monitored using CloudWatch. Event metadata can be stored in DynamoDB, while Lambda can support event-driven processing and orchestration.
We also plan to evaluate the Cloud-Optimized OpenCV Library (COOL) on AWS Graviton against an appropriate OpenCV baseline. The comparison will measure reproducible metrics such as latency, throughput, utilization, and processing cost.
Agentic Vision
MtaaGuard AI is intended for the Agentic Vision path because OpenCV 5 outputs directly influence later system behavior. A detected event does not simply generate a chatbot explanation. Instead, the visual result can trigger a new analysis step, change the inspection parameters, request more evidence, create an incident record, recommend an action, or ask for human approval.
Evaluation
We plan to measure detection precision and recall, tracking stability, event classification accuracy, false-alarm rate, end-to-end latency, throughput/FPS, and cloud processing cost. For the COOL path, we will document the deployment configuration and compare performance against a reproducible baseline.
Judge demo
The demonstration will start with a road video clip. The system will detect and track road users, build trajectories, identify a hazardous interaction, evaluate its severity, perform any required visual re-check, and display the resulting evidence and recommended action in the dashboard. We will also show benchmark results comparing the OpenCV baseline with the COOL/Graviton execution path.
Impact
MtaaGuard AI aims to make road-safety monitoring more proactive and accessible by extracting useful safety signals from ordinary cameras rather than requiring expensive new roadside infrastructure everywhere. The long-term goal is to help organizations identify recurring hazards, understand near misses, and intervene before those patterns become serious incidents.
Built With
- amazon-cloudwatch
- amazon-dynamodb
- amazon-web-services
- aws-graviton
- aws-lambda
- cool
- docker
- opencv-5
- python