MambaOut-CA Model for Construction Material Inventory Estimation Based on Ground Robot Panoramic Images

Abstract

In this work, we  develop a MambaOut-CA model for estimating building material inventories using ground-robot panoramic photos. The goal is to minimise the subjectivity and latency associated with hand counting in congested storage yards, mixed-material stacks, dust, fluctuating lighting, and shifting robot perspectives. Steel bars, cement bags, pipes, bricks, lumber, aggregate piles, formwork panels, and cable drums are used to create a simulated dataset of panoramic construction-site inspections. Build robot-view panoramas first, then utilise MambaOut sequence representation with channel attention to combine spatial, appearance, and inventory-state evidence after extracting material candidate locations and calibrating distance-sensitive scale features. In comparison to convolutional, transformer, and non-attentive fusion baselines, the suggested model reduced the counting mean absolute error to 3.7 items, lowered the volume estimation error to 6.4%, and achieved a category accuracy of 91.8% based on experiments with 14,200 panoramic frames and 58,600 annotated material regions. According to the findings, when material boundaries are absent or partially veiled, extended long-range panoramic context and channel-level selection enhance inventory estimate. The model provides a workable route for site-level replenishment planning, material scarcity alerts, and robot-assisted stock monitoring.