Logistics Hub Network and Edge AI Video Analytics

Challenge

Hub networks were built region by region; scanner or sorter disconnects immediately backed up parcels. Hundreds of cameras per hub were only reviewed after complaints, so rough handling, unsafe behaviour and blocked fire lanes went undetected in real time, and back-hauling all video to HQ was neither affordable nor practical.

GSS scope of responsibility

GSS owned deployment and 24/7 operation of site networks at 86 hubs, circuits to regional data centres, edge AI video-analytics nodes, and nationwide field engineering. Detection rules and response procedures were defined by the customer safety and operations teams.

Solution

GSS standardised all 86 hubs on dual-core, industrial-wireless site networks with dedicated circuits to regional data centres, deployed edge AI video-analytics nodes at every hub (rough handling, intrusion, fire-lane obstruction, forklift violations) that back-haul only alerts and clips, and provides nationwide field engineering with 7×24 managed operations.

Background

The carrier processes about 28 million parcels a day through automated sorters, PDA scanners and thousands of industrial devices at 86 hubs. Networks had been built regionally with single-point switches and patchy wireless; any outage stopped the sorter and parcels piled up — more than 400 network-caused backlog incidents in 2023.

Each hub has 300–600 cameras, used only to pull footage after a customer complaint. Damage claims from rough handling, safety incidents from staff entering sorter zones and goods stacked in fire lanes had no real-time detection. Management wanted AI video analytics, but the bandwidth and storage to stream every camera to HQ were prohibitive.

Solution

Hub network (traditional network business)

• Standardised site design: dual-core switching with ring-topology access at all 86 hubs; sorter PLCs, PDAs, cameras and office traffic in separate segments; industrial-grade APs covering sorting floors and loading docks so PDAs roam without drops.
• Circuits and backup: dual circuits from each hub to its regional data centre (primary MSTP + SD-WAN over internet backup) with second-level failover; 10G between the three regional data centres.
• Phased delivery: GSS field teams upgraded all 86 hubs in ten months, working in the 10:00–16:00 daily lull, with no sorting day affected.
• Operations: all hubs monitored 7×24 by the GSS NOC; spares pre-positioned at eight regional depots for four-hour on-site replacement.

Edge AI video analytics (AI business)

• Edge nodes: one or two GPU edge servers per hub ingest 300–600 local camera streams for real-time inference, sending only alert events and 15-second clips to HQ — under 1% of raw video bandwidth.
• Detection scenarios: rough handling (throwing, stepping on parcels), intrusion into sorter danger zones, fire-lane and emergency-exit obstruction, forklift speeding and passenger carrying, missing PPE.
• Model iteration: initial models trained on three months of labelled footage from two hubs; hub safety officers mark alerts as confirmed / false, and monthly model updates are pushed to all 86 hubs through the HQ cloud platform pipeline.
• Business loop: alerts go to the hub duty manager and the HQ safety dashboard; rough-handling events are automatically linked to workstation and shift for performance and training.
• Added value: the same streams feed sorter-congestion and dock-occupancy analytics into the dispatch system.

Results

Metric	Before	After
Network-caused backlog incidents	400+ in 2023	18 in 2025 to date (−95%)
PDA disconnect complaints	~1,200 / month	~60 / month
Rough handling detected	0 (post-hoc only)	~1,900 real-time alerts / day
Damage claims	Baseline	−31% over 6 months
Sorter-zone safety incidents	11 / year	1 since go-live
Video back-haul bandwidth	Infeasible (full streams)	< 1% of raw traffic

All hubs delivered by May 2025; GSS continues managed operations and monthly model iteration, and is replicating the design at 12 new hubs under construction.