Technical Summary: Coal-Gangue Detection and Pose Estimation Method Based on DRD-YOLOv11n

1. Problem Statement

The paper addresses the critical challenges in the intelligent sorting of coal and gangue (waste rock) within complex industrial environments. Traditional sorting methods (manual, dense medium separation) suffer from high labor intensity, low efficiency, and environmental hazards. While image recognition-based robotic sorting offers a solution, existing deep learning approaches face three specific bottlenecks when applied to on-site coal sorting:

1. Environmental Interference: Poor lighting, dust, and motion blur in underground mines degrade image quality, leading to feature extraction difficulties and edge loss.
2. Target Irregularity: Coal and gangue exhibit highly irregular shapes and large scale variations, which standard fixed-kernel convolutions struggle to localize accurately.
3. Grasping Limitations: Existing detection models typically output only bounding box coordinates. They fail to provide the pose parameters (specifically the rotation angle and geometric center) required for robotic grippers to align with irregular, elongated objects. This lack of pose information leads to low grasping success rates and high risks of mechanical collision in dense material streams.

2. Methodology

The authors propose DRD-YOLOv11n, a lightweight, one-stage detection and pose estimation framework based on the YOLOv11n architecture. The method integrates four key technical modules to address the aforementioned challenges:

A. Dynamic Upsampling via DySample

To mitigate the "checkerboard effect" and information loss caused by traditional nearest-neighbor interpolation (especially under low-light and motion-blur conditions), the feature fusion network replaces static interpolation with the DySample operator.

• Mechanism: DySample treats upsampling as a content-aware resampling process. It generates dynamic offsets based on input feature maps to adaptively adjust sampling point positions.
• Benefit: This allows the model to recover edge details and texture features at a low computational cost without relying on heavy dynamic convolutions.

B. Residual Depthwise Feature Enhancement (RD-SE)

To suppress background noise (dust, light variations) and enhance subtle texture differences between coal and gangue, a Residual Depthwise Feature Enhancement (RD-SE) module is embedded before the detection head.

• Structure: It combines a Residual Depthwise convolution unit (using 3×3 depthwise and 1×1 pointwise convolutions) with a Squeeze-and-Excitation (SE) attention mechanism.
• Function: The depthwise convolution preserves channel independence and reduces parameters, while the SE module recalibrates channel weights to emphasize key texture features and suppress noise.

C. Dynamic Detection Head based on DCNv3

To handle the irregular shapes and large scale spans of coal/gangue, the standard detection head is reconstructed into a DynamicDCNv3Head.

• Mechanism: This head integrates a unified attention mechanism with the DCNv3 (Deformable Convolution Network v3) operator. It employs three sequential attention mechanisms:
  1. Scale-Aware: Dynamically fuses features across different pyramid levels.
  2. Spatial-Aware: Uses DCNv3 to learn geometric deformations, allowing sampling grids to adapt to irregular boundaries.
  3. Task-Aware: Dynamically reweights channels to balance classification and regression tasks.
• Benefit: This significantly improves localization accuracy for targets with complex geometries and reduces background noise intrusion.

D. Pose Estimation and Anti-Collision Strategy

Beyond detection, the system includes a post-processing module to enable robotic grasping:

1. Pose Calculation: It extracts the target region, performs adaptive binarization (Otsu's method), and fits a Minimum Area Bounding Rectangle (MABR). The rotation angle is normalized and smoothed using Exponential Moving Average (EMA) filtering to prevent jitter.
2. Dynamic Safety Zone: To prevent collisions in dense stacks, the system constructs a dynamic "safe zone" around the target based on the gripper's physical dimensions and the target's pose.
3. Collision Avoidance: Using the Separating Axis Theorem (SAT), the system checks for intersections between the target's safe zone and other objects. If a collision risk is detected, the target is marked as ungraspable, ensuring the robot only attempts safe grasps.

3. Key Contributions

• Architectural Innovation: The integration of DySample, RD-SE, and DCNv3-based dynamic heads into the YOLOv11n framework specifically tailored for the harsh conditions of coal mining.
• Pose-Aware Detection: The development of a pipeline that transitions from simple object detection to full pose estimation (position + rotation), directly addressing the needs of robotic manipulation.
• Safety Mechanism: The implementation of a geometric constraint-based anti-interference strategy that proactively avoids collisions, a critical requirement for deploying robots in unstructured, dense environments.
• Lightweight Efficiency: Achieving high performance while maintaining a compact model size suitable for edge deployment.

4. Experimental Results

The method was evaluated on a custom dataset of 25,818 images (collected from the Dongrong No. 3 Coal Mine and a simulated lab environment) and compared against models like Faster R-CNN, SSD, RT-DETR, and various YOLO versions (v8, v11, v13).

• Detection Performance: The DRD-YOLOv11n achieved an mAP@0.5-0.95 of 85.6% and an F1 score of 95.1%. This represents a 0.6% improvement over the baseline YOLOv11n.
• Efficiency: The model size was reduced by 6% (down to 2.43M parameters) compared to the baseline, with a model weight of only 5.1 MB.
• Speed: The inference speed reached 114 FPS, outperforming YOLOv10n (86 FPS) and YOLOv13n (53 FPS).
• Grasping Success: In robotic grasping tests on elongated objects, the inclusion of pose estimation increased the success rate from 60% to 90%.
• Safety: In simulated dense stacking conditions, the anti-collision strategy achieved 100% collision avoidance and a 0% physical accident rate.

5. Significance and Claims

The paper claims that DRD-YOLOv11n provides a balanced solution that simultaneously optimizes detection accuracy, inference speed, and model complexity. By solving the "low positioning accuracy" problem through pose estimation and the "collision risk" problem through dynamic safety zones, the method offers a practical engineering reference for transitioning intelligent coal-gangue sorting robots from laboratory prototypes to industrial deployment.

The authors emphasize that their approach does not merely improve detection metrics but directly enables robust physical interaction (grasping) in complex, unstructured environments. The integration of lightweight modules (DySample, RD-SE, DCNv3) demonstrates that high-performance, real-time sorting is achievable on edge devices without the computational overhead of heavier two-stage algorithms.