研究目的
To propose a novel method for scene and position recognition based on visual landmarks (VLs) for autonomous mobile robots in human environments, focusing on robustness against human interference and environmental changes.
研究成果
The proposed method achieves higher recognition accuracy (49.9%) compared to a method without HOG masking (46.7%), demonstrating robustness to human interference. False recognition occurs mainly in neighboring zones, which can be reduced with better zone separation. Future work includes expanding datasets, incorporating incremental learning, and applying to larger environments.
研究不足
The method is sensitive to parameter settings, has computational costs (approx. 3s per image), and recognition accuracy is affected by environmental changes and human presence. Evaluation is limited to specific indoor environments and zones, with no comparison to state-of-the-art methods like deep learning.
1：Experimental Design and Method Selection:
The method uses saliency maps (SMs) to detect conspicuous regions, histograms of oriented gradients (HOG) for human region masking, and accelerated KAZE (AKAZE) for feature description. Adaptive category mapping networks (ACMNs) with self-organizing maps (SOMs) and counter propagation networks (CPNs) are used for recognition.
2：Sample Selection and Data Sources:
Benchmark datasets include KTH-IDOL2 for indoor scenes without humans, EPFL for human detection, and original datasets created using a mobile robot in a corridor divided into five zones, with and without humans.
3：List of Experimental Equipment and Materials:
A mobile robot (Double Robotics, Inc.), a monocular camera (built-in tablet camera), and a laptop computer for data processing.
4：Experimental Procedures and Operational Workflow:
Images are captured, converted to static images, processed to extract features using SMs, HOG masking, and AKAZE, and then classified using machine learning with LOOCV evaluation.
5：Data Analysis Methods:
Recognition accuracy is calculated using LOOCV, with confusion matrices for error analysis, and parameters optimized for SOMs and CPNs.
独家科研数据包，助您复现前沿成果，加速创新突破
获取完整内容
