Comparative Analysis of Texture Features and Deep Learning Method for Real-time Indoor Object Recognition

摘要

Object recognition and classification are considered as major tasks in the field of computer vision. They are well suited for applications such as a real-time system for people counting, object recognition system for people with visual impairments, surveillance systems, etc. The deployment of computer vision, machine learning, and deep learning algorithms enable to recognize the objects from an image or video frame. This paper proposes a real-time system for indoor object recognition. Moreover, the proposed work mainly focuses on analyzing the performance of various texture features, machine learning classifiers and deep learning methodologies to recognize the objects in indoor areas. The proposed methodology is validated in a publically available indoor object dataset “MCindoor20000”. The dataset consists of three categories of objects including doors, stairs, and sign. Our developed deep learning model using transfer learning approach yielded 100 % accuracy and texture features such as LPQ and BSIF have yielded an accuracy of more than 98% with SVM and KNN classifiers.