A multi-task X3D-based robotic video perception and state estimation framework for vehicle-loading counting in intelligent industrial systems

Abstract

Accurate recognition and counting of fully loaded vehicles in silo-based bulk-material loading scenes are essential for production management, logistics scheduling, and situational awareness in intelligent industrial systems. However, complex vehicle dispatching behaviors, severe visual contamination, and long-range temporal dependencies make this task difficult for conventional object detection or direct counting methods. To address these challenges, we formulate the task as robotic video perception and operational-state estimation rather than direct counting. The proposed framework decomposes industrial loading workflows into interpretable semantic states and converts noisy video observations into persistent, action-relevant loading events through deterministic state-transition reasoning. To improve robustness under industrial conditions such as dust interference, illumination variation, and occlusion, auxiliary object detection tasks for vehicles and loading silos are introduced to guide the model toward semantically effective regions. Based on this formulation, we design a lightweight multi-task video perception framework built upon the X3D architecture and integrate it with a finite-state-machine-based counting strategy. The resulting state representation can support higher-level scheduling, supervisory-control, and safety-monitoring modules without requiring them to reason directly over raw video streams. Experimental results demonstrate that the proposed method achieves reliable and robust counting performance in real industrial loading environments.