SWM Environment · Proposal recap

Edge AI Camera System for HSE Compliance Monitoring

A practical edge-computing pilot designed to monitor safety compliance and selected HSE events directly from CCTV video.
Tri Wind Technology Sdn Bhd
CAMERA
EDGE AI
MQTT
CLOUD
01 / 10
01 · Objective
Automate monitoring of key HSE compliance indicators and safety events from existing CCTV — with AI processing at the edge and alerts delivered to the cloud.
Monitor continuously
Use video analytics to detect defined compliance indicators and selected HSE events.
Reduce manual effort
Support safety teams with automated monitoring rather than depending on continuous human observation.
Keep processing local
Run inference on-site so AI processing remains available even when connectivity is limited.
02 / 10
02 · Pilot scope

Pilot at a glance

4
Cameras

Initial video sources for the pilot deployment.

1
Jetson edge device

Local compute for AI inference and event processing.

24/7
Local inference

Processing runs at the edge rather than relying on cloud inference.

MQTT
Cloud events

Detection events are transmitted to the client cloud backend.

CCTV
→
EDGE AI
→
EVENT
→
MQTT
→
SWM CLOUD
03 / 10
03 · HSE monitoring

What the system monitors

01
PPE
Helmet compliance
Detect helmet-wearing compliance and no-helmet cases.
02
VEST
Safety vest
Monitor safety-vest / uniform compliance in defined work areas.
03
GLOVE
Gloves / PPE
Support additional PPE compliance classes defined during annotation and model development.
04
NO
SMOKE
No-smoking zones
Detect prohibited smoking-related actions within monitored zones.
05
SPILL
Spill events
Monitor for spill conditions as part of the HSE event scope.
06
FIRE
Fire events
Monitor for fire-related visual events within the proposed system scope.
04 / 10
04 · System architecture

How it works

Input
CCTV / RTSP
Video streams from connected IP cameras.
→
Ingest
GStreamer
Hardware-accelerated video decoding and stream handling.
→
Inference
TensorRT AI
Optimized models run locally on NVIDIA Jetson.
→
Logic
Alert event
Detection logic turns model output into system events.
→
Cloud
MQTT
Events are published to SWM's cloud backend for handling.
Edge-first by design: AI processing operates locally for resilience in low-connectivity environments; online connectivity is primarily required for alert-event transmission.
05 / 10
05 · Solution scope

Four parts of the solution

01
Vision AI model
Annotation protocol, data labeling, synthetic-data setup, model training, versioning and TensorRT optimization.
02
Edge application
Jetson staging, containerized GStreamer / TensorRT processing, multi-camera ingest, MQTT alerts and performance monitoring.
03
Monitoring dashboard
Web dashboard for live video preview and operational metrics, with Admin and Viewer access roles.
04
OTA deployment
K3s-based orchestration, CI/CD deployment and rollback, plus documentation and training for SWM IT.
06 / 10
06 · AI development

From data to deployed model

01
Annotation protocol
Define classes and ambiguous cases.
→
02
Label data
Label client-provided video frames.
→
03
Synthetic data
Supplement rare negative-event examples.
→
04
Train & version
Train cascaded models and version datasets/models.
→
05
TensorRT
Optimize models to FP16 TensorRT engines for edge inference.
70–80%
Initial pilot accuracy target stated in the proposal. Evaluation framework to be defined after project initiation.
07 / 10
07 · Delivery plan

Project phases & deliverables

4 weeks
Data preparation
Annotation protocol, labeled dataset, synthetic-data setup.
6 weeks
Model training
Trained and optimized TensorRT AI model.
5 weeks
Edge application
Containerized GStreamer + MQTT alert pipeline.
4 weeks
Dashboard
Custom Phoenix dashboard with RBAC and live preview.
3 weeks
Integration & handover
Full pilot deployment, OTA setup, documentation and training.
Durations shown are the phase durations stated in the submitted proposal. The proposal does not define an overall critical-path calendar or explicit phase overlap.
08 / 10
08 · Commercial & handover

Commercial summary

Software project · with source code
RM198K
Before SST · after RM32K gratitude discount
Project subtotal incl. 8% SSTRM213,840
CCTV installation servicesRM86,420 + SST
Grand total shown in proposalRM307,173.60
Ownership & handover
  • Developed software, source code, datasets and model weights handed over after completion and full payment.
  • SWM receives access and ownership for internal maintenance and retraining.
  • 30 days of warranty support for bug fixes and deployment stability.
  • SWM IT assumes ongoing management after handover.
09 / 10
09 · Meeting discussion

Discussion & next steps

01
Confirm pilot use cases
Align on the compliance and HSE events to prioritize during the four-camera pilot.
02
Finalize hardware
Confirm Jetson and camera specifications during the system-design phase based on performance and cost.
03
Prepare data & feedback
Coordinate video-data delivery, labeling inputs and feedback cycles needed for model development.
04
Define evaluation & kickoff
Establish the evaluation framework after project initiation and align on implementation coordination with SWM.
Objective for this meeting: align on scope, clarify questions, and agree the path to pilot kickoff.
10 / 10
01 / 10