One Touch - AI-Powered Computer Vision And Safety Operations Platform

About The Client

One Touch is a multi-tenant AI computer vision and safety operations platform designed to turn CCTV, edge-camera, thermal, and uploaded video streams into structured detections, real-time alerts, evidence, operational cases, and management insights.

Bitontree's AI team was responsible for the core vision intelligence behind the platform: dataset development, annotation strategy, preprocessing, YOLO model training and fine-tuning, transformer-based visual classification, temporal behavior understanding, evaluation, optimization, and production inference. We also helped define the common AI service architecture that allows up to 20 models to run independently or in parallel for a single video job.

The platform connects AI detections with a broader operating environment that includes a Security Operations Center (SOC), Workplace Safety Operations (WSO), Enforcement and Case Management (ECM), an Operations Support Center (OSC), device management, evidence handling, tenant administration, and dealer/customer portals.

Project Detail	Information
Client	One Touch
Public Client Name	TODO: Confirm brand styling for public use
Industry	Physical Security, Workplace Safety, Computer Vision SaaS
Location	TODO: Confirm client location
Duration	TODO: Confirm duration
Project Dates	TODO: Confirm project dates
Project Scope	Computer vision model development, optimized inference, and multi-model AI service architecture
Services Used	AI Model Development, Computer Vision Engineering, MLOps, AI Agent Development, Custom Software Development, Cloud Engineering, Edge Deployment
Engagement	AI intelligence layer for a wider safety and operations platform
Live Link	TODO: Confirm public link
Client Contact / Testimonial	TODO: Confirm if a testimonial or client quote can be included

---

The Objective

One Touch wanted to turn existing camera infrastructure into an active safety, security, and operations system rather than a passive recording estate.

The goal was a production AI platform, not a set of standalone detection demos. Detections had to be accurate under real site conditions, reach the teams who act on them, and improve over time from operator feedback.

The platform needed to:

• Cover the full computer vision lifecycle from data collection through annotation, training, evaluation, deployment, monitoring, and retraining
• Detect people, PPE, phones, vehicles, plates, fire, smoke, hazards, and other visual classes across varied cameras and lighting
• Understand temporal events such as aggression, falls, sleep, idling, and unauthorized grouping
• Run up to 20 models independently or in parallel against a single video job
• Meet real inference budgets on cloud GPUs, on-premise servers, and edge appliances
• Keep tenants, media, and model access isolated on shared infrastructure
• Convert detections into alerts, evidence, cases, dashboards, and auditable operator actions

---

The Challenge

One Touch needed more than isolated computer vision demos. It required a production-ready AI platform capable of handling multiple safety, security, workforce, vehicle, and operational use cases across different tenants, sites, cameras, and environmental conditions.

The main challenges included:

• Building reliable, site-specific datasets from imperfect real-world video
• Training and fine-tuning detection models for small, occluded, low-light, and visually similar classes
• Supporting both frame-level object detection and temporal behavior understanding
• Running multiple AI models against the same video without creating a tightly coupled monolith
• Optimizing models for GPU and edge deployment using ONNX Runtime and NVIDIA TensorRT
• Preserving tenant isolation while serving many customers and devices from a shared platform
• Turning detections into operational alerts, evidence clips, cases, dashboards, and auditable actions
• Reducing false positives through hard-negative mining, temporal validation, and human feedback
• Maintaining model versioning, reproducibility, performance metrics, and a path for automated retraining

The solution therefore had to cover the complete computer vision lifecycle, not only inference: data collection, cleaning, annotation, augmentation, training, fine-tuning, evaluation, deployment, monitoring, feedback, and continuous improvement.

---

Our Approach

Bitontree designed the AI layer as a modular model platform with a shared preprocessing pipeline, standardized model interface, independent model workers, parallel job orchestration, and a unified inference-output schema.

Our AI engineering approach focused on:

• Developing and curating datasets for each camera environment and detection objective
• Building preprocessing, augmentation, class-balancing, frame-selection, and filtering workflows
• Using YOLO as the primary object-detection family for people, PPE, vehicles, phones, plates, fire, smoke, hazards, and other visual classes
• Fine-tuning models on customer-specific footage to improve real-world performance
• Combining detection, tracking, pose estimation, OCR, geometric rules, and temporal logic where one model alone was insufficient
• Using transformer-based models for Make-Model-Color (MMC) vehicle recognition
• Using VideoMAE-based temporal video understanding for aggression and complex activity recognition
• Exporting trained models to ONNX and accelerating inference with ONNX Runtime and TensorRT
• Creating standardized APIs and worker services so models could be deployed, replaced, scaled, and monitored independently
• Connecting detections to alerts, evidence, case workflows, dashboards, and human review

---

Core Features Developed

End-to-End Dataset Engineering

We built the dataset foundation required for production computer vision. This included cleaning corrupted, blurry, obstructed, irrelevant, or redundant frames; selecting useful frames from video; normalizing inputs; enhancing low-light imagery; reducing noise; applying geometric transformations; and handling difficult backgrounds.

We created task-specific annotation strategies for bounding boxes, polygons, segmentation masks, identity labels, OCR text, human keypoints, zones, and temporal action classes. Dataset augmentation and automatic class balancing were used to improve generalization and reduce bias toward overrepresented classes.

Training datasets, checkpoints, metrics, confusion matrices, reports, and evaluation artifacts were stored with versioned paths so experiments could be reproduced and audited.

YOLO Model Training And Fine-Tuning

YOLO-based models formed the core of the detection stack. Our AI team trained and fine-tuned models on target-environment data rather than relying only on generic pretrained weights.

The workflow covered:

• Dataset construction and class definition
• Annotation and quality review
• Transfer learning from pretrained YOLO weights
• Hyperparameter tuning and augmentation
• Precision, recall, F1, IoU, and mAP evaluation
• Per-class error analysis and false-positive categorization
• Hard-negative mining using normal site footage
• Site-specific fine-tuning for camera angle, lighting, distance, uniforms, equipment, and background conditions
• Model checkpointing, versioning, and regression testing

This approach supported rapid development of new detectors while preserving a consistent evaluation and deployment process.

Optimized Inference With ONNX Runtime And TensorRT

We created a deployment path from trained PyTorch/YOLO models to optimized inference artifacts.

Models were exported to ONNX for portable, hardware-aware inference with ONNX Runtime. On NVIDIA GPU systems and edge appliances, TensorRT optimization was used to reduce latency and improve throughput through graph optimization, kernel fusion, precision tuning, and engine-level acceleration.

This allowed the same AI capability to be deployed across cloud GPU workers, on-premise servers, and supported edge hardware while keeping model inputs and outputs consistent.

Multi-Model Parallel Processing

The platform supports up to 20 computer vision models operating as isolated workers. A single API job can select one model or multiple models, and the dispatcher creates parallel sub-jobs rather than running every model sequentially.

Each model has its own queue and worker pool, loads its weights once at startup, processes jobs statelessly, emits health heartbeats, and returns a standardized result. A result aggregator merges detections, metadata, processing times, annotated media, and partial-failure information into one response.

Common AI Model Interface

We defined a common model lifecycle across all services: load, preprocess, run, postprocess, and execute.

This made YOLO detectors, classifiers, OCR pipelines, pose models, and transformer-based video models interchangeable within the orchestration layer. Models could be hot-swapped or upgraded without rewriting the API gateway or client applications.

Automated Evaluation And Continuous Improvement

The evaluation pipeline loads ground truth and predictions, converts coordinates, applies IoU-based matching, separates false-positive types, produces image-level visual feedback, and exports structured evaluation reports.

A feedback database links operator corrections to the original media. Confirmed false positives can be converted into hard-negative datasets, and retraining can be triggered after a configurable number of reviewed corrections. This creates a practical human-in-the-loop improvement cycle.

Person Detection, Attendance, And Zone Presence

We implemented person detection with single-camera multi-object tracking and persistent track IDs. Detected people can be mapped to entry, exit, work, and restricted-zone polygons to measure presence, loitering, attendance events, and time spent in an area.

Identity can be layered through face recognition, badge OCR, and uniform or role classification, with recognition priority determined by the available visual evidence.

Worker Identification

The worker-identification pipeline combines multiple signals:

• RetinaFace-style face detection and ArcFace-style embeddings
• Badge detection and OCR for employee numbers
• Uniform and role classification
• Enrollment and vector search using an embedding database such as FAISS

The layered approach improves operational matching while acknowledging camera resolution, occlusion, privacy, and placement constraints.

PPE Detection And Compliance Mapping

YOLO models were trained and fine-tuned to detect people and required PPE classes such as hard hats, safety vests, gloves, and safety boots.

Post-processing maps PPE detections to the correct person using geometric relationships and IoU rules. Violations generate structured evidence containing the snapshot, bounding boxes, missing item, timestamp, zone, and worker identity when available.

Mobile Phone Usage Detection

The phone-usage pipeline detects people and visible mobile phones, associates the phone with the relevant person, applies zone and policy rules, and generates an alert when usage occurs in a restricted work area.

Sleep And Drowsiness Detection

We built a custom training workflow for awake and sleeping states using annotated site footage. The solution combines person localization with posture, head position, and temporal persistence so a momentary pose is not automatically treated as sleep.

Fall Detection

Fall detection combines person detection, pose estimation, body orientation, state transitions, and temporal validation. Models are trained and fine-tuned using fall, standing, sitting, and non-fall examples, with alerts generated only after the post-processing logic validates the event across frames.

Aggressive Behaviour Detection With VideoMAE

Aggression cannot be understood reliably from a single frame. We used a temporal pipeline that combines person detection, optional pose/keypoint features, and a VideoMAE-based transformer model fine-tuned on video clips labelled as aggressive or normal.

The model analyzes motion across time to identify patterns such as fighting, pushing, punching, or threatening body movement. Temporal confidence, persistence rules, and event clipping are used to reduce false alerts and preserve the evidence surrounding the incident.

Unauthorized Grouping And Crowd Analytics

People detection and tracking are combined with zone polygons, clustering, occupancy thresholds, density estimation, and temporal flow analysis to identify unauthorized gatherings or overcrowding.

The platform can produce people counts, density scores, heatmaps, entry/exit trends, dwell time, congestion signals, and alerts when a configured group threshold is exceeded in a restricted area.

Inactivity And Idling Detection

Tracked workers are monitored over time to measure displacement, dwell duration, and sustained lack of movement. When inactivity exceeds a configurable policy threshold, the system creates an alert with the worker details, start time, zone, and snapshot.

Make-Model-Color Recognition With Transformers

For vehicle analytics, we used a transformer-based visual model for Make-Model-Color (MMC) recognition. Vehicle crops produced by the detector are normalized and passed to the MMC classifier to infer manufacturer, model family, and color attributes.

This enriches number-plate results and supports vehicle search, mismatch analysis, enforcement workflows, and investigation even when a plate is partially unreadable.

ANPR And Vehicle Intelligence

The ANPR pipeline covers plate detection, four-corner annotation, perspective rectification, contrast enhancement, OCR, regional syntax validation, and character-level post-processing.

Vehicle intelligence can also include vehicle-type detection and tracking, MMC classification, registration lookup, expired-registration checks, and wanted or blacklist matching.

Fire And Smoke Detection

We trained YOLO-based fire and smoke detectors using real and synthetic data, segmentation-oriented annotations where appropriate, and hard examples such as sun glare, steam, welding sparks, and reflections.

HSV-based color validation, consecutive-frame checks, and hard-negative mining help reduce nuisance alerts in real operating environments.

Restricted Zone And Virtual-Fence Detection

Operators can define zone polygons and virtual fences for fixed camera views. Person, vehicle, or machinery detections are evaluated against these boundaries, tenant policies, schedules, and authorization rules before a breach alert is generated.

Unsafe Behaviour And Sequence Analysis

Temporal video models and rule-based post-processing support visually defined unsafe actions such as running, climbing, jumping, unsafe access, and other customer-specific movement patterns.

The architecture also supports historic behavior matching, violation retrieval, and sequence similarity for investigations, while keeping human review available for subjective or low-confidence cases.

Workplace Hazard Detection

Custom detectors identify visible trip hazards such as boxes and obstacles in configured walkway zones. Persistence timers filter moving objects, and validation against empty walkways helps reduce false positives caused by lighting and shadows.

Thermal And Heat Monitoring

Thermal models are trained on thermal-specific datasets rather than assuming RGB pretrained models will transfer directly. The detector first localizes the target or region of interest, after which pixel-intensity or calibrated temperature logic determines normal and high-temperature states.

Object Tracking And Evidence Extraction

DeepSORT or ByteTrack-style tracking assigns persistent IDs to people and vehicles across frames. When an incident is confirmed, the evidence service extracts the exact snapshot and a configurable video window around the timestamp, applies enhancement or ROI cropping, and links the media to the alert or case.

AI APIs, Model Registry, And Deployment Management

The platform exposes APIs for data ingestion, dataset management, annotation, preprocessing, training control, model testing, inference, feedback, model registry access, deployment, monitoring, and logs.

Models can be packaged as Dockerized microservices and deployed to cloud, on-premise, or edge environments. Per-tenant thresholds, model versions, sampling rates, and output settings are externalized rather than hardcoded.

Security Operations Center

AI alerts feed a live SOC wall with real-time tiles, severity grouping, device and ANPR overlays, quick actions, and sub-second update targets. The triage console supports priority scoring, alert details, snapshots or streams, acknowledgment, classification, assignment, escalation, resolution, notes, and audit history.

Workplace Safety Operations

The WSO experience consolidates PPE, mobile usage, inactivity, sleep, aggression, unsafe behavior, grouping, fall, and danger-zone alerts. Supervisors can review behavior KPIs, worker profiles, PPE requirements, safety violation history, and risk or compliance scores.

Enforcement And Case Management

Confirmed AI detections can automatically create vehicle or worker cases. Case workflows include evidence viewing, annotations, classification, notes, related events, registry-versus-camera comparison, report export, authority handoff, supervisor email, blacklist actions, and closure.

Evidence Locker And Chain Of Custody

Images and video clips are stored with tenant, site, device, timestamp, alert type, model version, and case metadata. Search, filtering, export, access logging, encryption, retention policies, legal hold, and immutable chain-of-custody records support investigation and compliance requirements.

Device And Operations Monitoring

The OSC monitors connectivity, latency, packet loss, throughput, media delivery, AI detection frequency, model reliability, SIM usage, signal quality, storage, CPU, memory, firmware drift, and overall device or site health.

Operators can run connectivity, snapshot, stream, configuration, and log diagnostics; create faults and work orders; assign technicians; escalate to vendors; and perform permitted remote operations.

Multi-Tenant Administration

The wider platform supports dealer and customer onboarding, platform-to-dealer-to-customer-to-site-to-zone hierarchy, tenant isolation, module licensing, detection profiles, alert-routing rules, device permissions, feature flags, SSO, SCIM, MFA, API keys, secrets, webhooks, retention, backup, legal hold, branding, RBAC, and complete admin audit trails.

---

Technology Stack

Layer	Technology
Model Development	Python, PyTorch, Ultralytics YOLO, OpenCV
Temporal Video AI	VideoMAE, transformer-based action recognition, pose estimation
Vehicle Intelligence	Transformer-based MMC classification, YOLO, OCR
Optimized Inference	ONNX, ONNX Runtime, NVIDIA TensorRT, CUDA
Tracking	ByteTrack / DeepSORT-style multi-object tracking
Face & Identity	RetinaFace-style detection, ArcFace-style embeddings, FAISS
OCR	PaddleOCR / CRNN / LPRNet-style pipelines
API & Orchestration	FastAPI, Pydantic, asyncio, API Gateway, worker pools
Queue & Cache	Redis Streams or RabbitMQ, Redis
Database	PostgreSQL, SQLAlchemy, Alembic
Storage	AWS S3-compatible object storage, tenant-prefixed media storage
Containerization	Docker, Docker Compose, GPU-enabled containers
Monitoring	Prometheus, Grafana, structured JSON logging, ELK-compatible logs
Model Registry	Versioned S3/DVC-style model and dataset registry
Security	JWT, API keys, RBAC, TLS, encrypted evidence, tenant isolation

---

Product Architecture

System	Technology / Pattern	Role
Client Applications	Web dashboards and operational portals	Submit jobs, configure models, review alerts, manage cases, devices, and tenants
API Gateway	FastAPI, JWT/API keys, rate limiting	Authenticate tenants, validate payloads, enforce subscriptions, and route model jobs
Job Dispatcher	Async service with priority and fairness controls	Split multi-model requests into parallel sub-jobs and manage lifecycle state
Queue Manager	Redis Streams / RabbitMQ	Maintain dedicated model queues, retries, timeouts, and dead-letter handling
AI Worker Pool	PyTorch, ONNX Runtime, TensorRT	Run isolated detection, classification, tracking, OCR, and temporal video models
Result Aggregator	Standardized inference schema	Merge model outputs, processing metadata, detections, and annotated media
Evidence Service	S3-compatible storage, encryption, audit logs	Store snapshots and clips with retention and chain-of-custody controls
SOC / WSO / ECM	Real-time event and case workflows	Turn AI detections into triage, safety, enforcement, and investigation actions
OSC / Device Management	Telemetry, diagnostics, configuration services	Monitor cameras, edge devices, streams, firmware, resources, and AI health
Tenant Administration	PostgreSQL, RBAC, configuration services	Manage subscriptions, modules, users, sites, zones, branding, and governance

---

The Result

Bitontree's AI team delivered the model-development foundation and production inference architecture for a broad computer vision SaaS platform.

Outcome	Result
Dataset development	Repeatable cleaning, annotation, augmentation, balancing, frame-selection, and artifact-versioning workflows
YOLO engineering	Custom detectors trained and fine-tuned for people, PPE, phones, vehicles, plates, fire, smoke, hazards, and other classes
Temporal intelligence	VideoMAE-based aggression recognition and temporal logic for behavior, sleep, falls, inactivity, and grouping
Vehicle intelligence	ANPR, tracking, vehicle classification, and transformer-based MMC recognition
Deployment performance	Portable ONNX models with ONNX Runtime and TensorRT acceleration for server and edge inference
Parallel processing	One video job can trigger multiple isolated models and receive one aggregated result
Continuous improvement	Human feedback, hard-negative mining, evaluation reports, and retraining triggers connected to the data lifecycle
Operationalization	AI detections converted into alerts, evidence, dashboards, cases, blacklists, reports, and audit trails
Multi-tenancy	Tenant-aware authentication, model access, rate limits, queues, databases, and storage namespaces
Platform observability	Model latency, queue depth, detection health, device health, failures, retries, and usage made measurable

---

Business Impact

The platform transforms passive video infrastructure into an active safety, security, compliance, and operational intelligence system.

The solution helps customers:

• Detect safety violations and behavioral risks earlier
• Reduce manual monitoring across large camera estates
• Turn AI events into evidence-backed operational workflows
• Improve PPE compliance and workplace accountability
• Identify phone use, sleep, falls, aggression, idling, and unauthorized grouping
• Strengthen vehicle monitoring with ANPR, tracking, and MMC recognition
• Run many models through one consistent API and deployment architecture
• Deploy optimized inference closer to cameras for lower latency
• Improve accuracy over time using reviewed false positives and site-specific retraining
• Keep customer data and media separated within a multi-tenant platform
• Monitor both AI performance and the health of the devices delivering the video
• Preserve auditable evidence and operator actions for compliance and investigations

---

Key Engineering Problems We Solved

Building Datasets From Real Camera Conditions

Production footage contains blur, occlusion, compression, reflections, changing light, difficult angles, and long periods without useful events. We created preprocessing, frame-selection, annotation, augmentation, and quality-control workflows that convert this raw footage into trainable and auditable datasets.

Fine-Tuning YOLO For Site-Specific Accuracy

Generic object detectors do not automatically understand a customer's PPE, uniforms, camera geometry, plate formats, machinery, or hazard definitions. We fine-tuned YOLO models on target footage and used class balancing, hard negatives, and per-class error analysis to improve reliability.

Understanding Actions Across Time

Aggression, falls, sleep, idling, and grouping are temporal events rather than static objects. We combined detectors with tracking, pose, sequence logic, and VideoMAE-based transformer models to reason across multiple frames.

Accelerating Models For Edge And GPU Deployment

Research models must meet real inference budgets. We created an optimization path through ONNX Runtime and TensorRT so trained models could run efficiently on supported GPUs and edge systems without changing the client-facing response contract.

Combining Multiple Models Without Creating A Monolith

Each model has different compute, latency, data, and scaling requirements. Dedicated queues and workers allow slow or specialized models to operate independently while the aggregator still returns one coherent job result.

Reducing False Positives In Safety-Critical Alerts

We used IoU rules, zone logic, confidence thresholds, multi-frame persistence, pose validation, color checks, syntax checks, hard-negative mining, and human review to stop raw detections from becoming uncontrolled alerts.

Standardizing Diverse Model Outputs

Object detectors, classifiers, OCR engines, tracking systems, and temporal transformers produce different raw outputs. A common inference schema normalizes labels, confidence, timestamps, bounding boxes, attributes, processing time, errors, and evidence URLs.

Connecting AI To Operational Action

A detection becomes valuable only when it reaches the right workflow. We connected AI results to SOC triage, WSO compliance, enforcement cases, evidence capture, notifications, dashboards, blacklists, and audit history.

Maintaining Tenant Isolation Across The AI Pipeline

Tenant context is enforced across authentication, subscription-based model access, job queues, database queries, storage prefixes, compute quotas, results, logs, and evidence access.

Creating A Feedback-Driven Model Lifecycle

Operator decisions are mapped back to source media and model versions. Reviewed false positives and missed detections can become new training material, making continuous improvement part of the platform rather than a separate manual exercise.

---

Development Timeline

Phase	Deliverables
Phase 1	Platform architecture, repositories, Docker environment, database schema, queues, object storage, tenant controls, and common AI model interfaces
Phase 2	Dataset ingestion, cleaning, annotation, augmentation, class balancing, frame selection, artifact versioning, automated evaluation, visual diagnostics, and feedback workflows
Phase 3	YOLO training and fine-tuning for people, PPE, phones, vehicles, plates, zones, fire, smoke, hazards, and other classes; tracking, OCR, identity, pose estimation, fall, sleep, inactivity, and grouping logic
Phase 4	Transformer-based MMC and VideoMAE aggression/action models; ONNX export, ONNX Runtime integration, TensorRT acceleration, benchmarking, and GPU/edge optimization
Phase 5	Multi-model orchestration, worker pools, result aggregation, retries and DLQ handling; SOC, WSO, ECM, evidence, device-health, tenant and admin integrations; load testing, security validation, monitoring, deployment hardening, and production runbooks

---

Why This Project Stands Out

One Touch required a complete AI product capability, not a collection of pretrained models. The work spans dataset engineering, YOLO training and fine-tuning, temporal transformers, vehicle MMC classification, VideoMAE aggression recognition, model evaluation, ONNX Runtime and TensorRT optimization, parallel orchestration, multi-tenant security, edge deployment, and continuous feedback.

Bitontree's AI team built the intelligence layer that makes the wider safety and operations platform possible. The result is a modular computer vision foundation where new models can be trained, optimized, deployed, monitored, and connected to real business workflows without rebuilding the entire system.