Remote intelligent perception system for multi-object detection.

Area of Science:

• Computer Vision
• Robotics
• Artificial Intelligence

Background:

• Recent advancements in visual sensor technology have increased interest in classifying robotic environment scenes.
• Object detection and scene understanding are crucial for augmented reality, robot navigation, autonomous driving, and tourism applications.
• Challenges include semantic understanding, occlusion, data scarcity, illumination variations, and object scale differences.

Purpose of the Study:

• To propose an innovative scene recognition framework to address existing challenges in robotic environment image analysis.
• To enhance the accuracy and robustness of scene classification for diverse robotic applications.

Main Methods:

• Preprocessing using kernel convolution.
• Semantic segmentation via UNet.
• Feature extraction using discrete wavelet transform (DWT), Sobel, Laplacian, and local binary pattern analysis.
• Object recognition with deep belief networks and object-to-object relation analysis.
• Scene labeling using AlexNet.

Main Results:

• The framework achieved over 96% accuracy on the PASCALVOC-12 dataset.
• A 95.90% accuracy rate was obtained on the Cityscapes dataset.
• The model demonstrated 92.2% accuracy on the Caltech 101 dataset.

Conclusions:

• The proposed scene recognition framework is highly effective and yields remarkable results.
• The model represents a noteworthy advancement beyond current scene recognition capabilities.
• The framework shows significant potential for improving robotic perception and autonomous systems.