Dynamic and Real-Time Object Detection Based on Deep Learning for Home Service Robots.

Area of Science:

• Robotics
• Computer Vision
• Artificial Intelligence

Background:

• Home service robots require precise object identification and localization for efficient task execution.
• Motion-blurred and occluded images from mobile sensors pose significant challenges for object detection.
• Detecting small and occluded objects, common in daily life, is particularly difficult.

Purpose of the Study:

• To develop a dynamic and real-time object detection algorithm for home service robots.
• To address challenges posed by motion blur and object occlusion in indoor environments.
• To improve the accuracy and efficiency of object recognition for service robot applications.

Main Methods:

• Proposed a novel dynamic and real-time object detection algorithm comprising image deblurring and object detection components.
• Developed the DA-Multi-DCGAN algorithm for deblurring motion-blurred images using dynamic adjustment and multimodal fusion.
• Introduced the AT-LI-YOLO method for small and occluded object detection, incorporating attention mechanisms and a lightweight network structure.

Main Results:

• DA-Multi-DCGAN improved Peak Signal-to-Noise Ratio (PSNR) by 5.07 and Structural Similarity (SSIM) by 0.022 compared to DeblurGAN.
• AT-LI-YOLO achieved a 3.19% increase in mean average precision (mAP) over YOLOv3, with significant gains in detecting small (19.12%) and occluded (29.52%) objects.
• The integrated algorithm achieved a processing time of 29 ms using TensorRT, meeting real-time requirements for service robots.

Conclusions:

• The proposed dynamic and real-time object detection algorithm effectively handles motion blur and occlusion.
• The algorithm demonstrates superior performance in accuracy and efficiency for object detection in home service robot applications.
• The developed method enables smoother and more reliable operation of service robots in complex indoor environments.