Geospatial Processing and Computer Vision

Scalable 360° stitching and LiDAR-synced infrastructure inspection

Input
360° JPEG imagery with EXIF geo-tags, LiDAR LAS clouds, IMU telemetry, CSV assets, ESRI basemaps
Output
Map-synced stitched video with LiDAR overlays, browser viewer, geo-referenced defect pin-drops
Core Methods
FFmpeg stitching, overlap deduplication, tile segmentation, GPU burst processing, timeline synchronization, Three.js visualization, auto-scaling cloud compute
Deployment
S3 ingestion, Lambda-triggered scaling, short-burst GPU/high-core instances, browser delivery, regional fail-over, lifecycle storage

Problem and Solution

Problem: Scale and Cost

Massive 360° datasets required synchronized corridor inspection, but proprietary tools were costly, lacked LiDAR, and failed at scale.

Solution: Chunked Processing

Uploads were chunked, stitched with FFmpeg, and overlap duplicates removed for clean composites.

Problem: Accuracy & Drift

Early outputs showed duplicated objects and misalignment across imagery, LiDAR, and maps.

Solution: Unified Synchronization

LiDAR was tiled, denoised, and aligned with imagery and ESRI via a shared timeline.

Problem: Slow Delivery

Processing took days and large uploads were unreliable.

Solution: Scalable Rendering

Event-driven auto-scaling enabled minutes-level rendering with near-flat cost, delivered through a unified browser viewer.

Problem Definition

Inspection-grade synchronization of multi-modal corridor data at extreme scale

Visual task to be solved

Produce seamless 360° video aligned with ESRI while enabling synchronized LiDAR inspection across very large datasets.

Key technical constraints

  • Millions of frames and heavy overlaps
  • Different sampling rates across data streams
  • Multi-day single-thread runtimes
  • Browser memory limits for point clouds
  • Secure, reliable upload and recovery

Technical success criteria

  • Minutes-level processing
  • Reliable handling of hundreds-GB uploads
  • Improved duplicate detection via QA
  • High first-pass success with retry
  • Limited cost growth despite speed gains

Data Characteristics

High-volume, multi-sensor corridor capture with real-world edge cases

Data source and scale

Road-survey imagery from 10k frames to multi-million projects, plus LiDAR, telemetry, GIS basemaps, and CSV assets.

Labeling method

Algorithmic duplicate detection refined through visual QA review and operator feedback.

Primary data challenges

  • Overlap duplicates and repeated coordinates
  • Huge panoramas needing tiling
  • Near-identical corridor frames
  • Unstable field-network uploads
  • Browser limits for dense point clouds

Technical Approach

Chunk-first, event-driven architecture for synchronized large-scale processing

High-level pipeline design

Chunked S3 uploads → metadata and tiling → FFmpeg stitching → redundancy pruning → LiDAR decompression, denoising, tiling → master timeline sync → streamed browser visualization.

Key design decisions or trade-offs

  • Short bursts on large compute vs long small-instance runs
  • Fixed-size chunks for linear scaling
  • Ordinal alignment where GPS repeats
  • Prefetching instead of recompute
  • Object storage as source of truth

Non-obvious constraints handled

  • Fly-over/under-pass duplication
  • Memory-safe LiDAR tile streaming
  • Hash-chain frame integrity
  • Expiring role-scoped storage
  • Rapid regional fail-over

Models and Algorithms

Classical video processing with targeted detection and spatial methods

Model architectures or families

Similarity correlation, hashing, YOLOv8 symbol detection, Three.js rendering, tiled point-cloud visualization.

Core algorithms or methods

  • Histogram overlap pruning
  • Content-aware frame hashing
  • 4k tile segmentation
  • Voxel-grid LiDAR denoising
  • Quaternion-based synchronization

Training or optimization strategy

  • Multithreaded GPU-friendly batching
  • Priority-based auto-scaling
  • Cached FFmpeg builds
  • Human-in-loop QA feedback
  • Client-side WASM preview processing

Evaluation and Results

Minutes-level rendering with strong operational impact

Metrics used

Processing time, cost per batch, success rate, duplicate accuracy, viewer latency, decision time.

Quantified performance

~24h to ~5 min

10k images processing time

72h to ~8 min

12k images processing time

~100x throughput

for ~2x spend

97% success

on 500 GB uploads

~60% reduction

in storage cost

~85% faster

quote-to-decision cycle

Baseline comparison

1

Replaced costly proprietary tools lacking LiDAR and scale.

2

Removed manual duplicate correction and fragmented GIS workflows.

3

Delivered single-link browser inspection with synchronized imagery, LiDAR, and ESRI data.