PointTransformer: Encoding Human Local Features for Small Target Detection.

Main Methods:

• Introduction of Point Transformer, a transformer encoder, as the core backbone for object detection.
• Utilizing a priori information of human skeletal points to extract local features.
• Employing self-attention and cross-attention mechanisms to reconstruct local features for each key point.
• Proposing a learnable positional encoding method to enhance the model's sensitivity to skeletal point positions.

Main Results:

• The proposed Point Transformer model significantly outperforms classical object detection algorithms on a chemical plant field operation dataset.
• The model achieves a 12% improvement in mean average precision (mAP) for small target detection compared to state-of-the-art methods.
• Demonstrated superior performance in handling occlusion and detecting small targets in challenging industrial environments.

Conclusions:

• The Point Transformer framework effectively addresses the limitations of traditional object detection methods for small targets in occluded industrial settings.
• The integration of skeletal point information and learnable positional encoding significantly boosts detection accuracy.
• The developed model offers a promising solution for enhancing safety and operational efficiency in chemical plant field operations through improved surveillance.