Research On Object Detection Technology Of Mall Service Robot Based On Deep Learning

With the popularity of smart retailing in mall retailing,smart retailing can affect users' purchasing efficiency and user experience.Applying mall service robots is one of the solutions for smart retailing.Among them,the visual system is the difficulty and core of the service robot in the mall.Through the visual system to detect and analyze the pedestrians in the mall,it can complete the core functions of precise marketing and personalized services for pedestrians.Due to the high complexity of the ma ll environment and the thousands of pedestrians in the mall,there is currently a lack of detection and recognition technology for mall pedestrians with high detection accuracy and strong generalization capabilities.In the past,most of the detection mode ls applied to mobile robots used gradient histograms and support vector machines.The recognition models generally used supervised deep learning models as the basic classifiers.These methods have low detection accuracy and weak generalization ability.The Faster R-CNN target detection model and the zero-sample learning model can solve the above problems.The Faster R-CNN model can quickly and accurately detect the target.The zero-sample learning model is designed to give the detector the ability to recogn ize unknown categories,so that the model generalization ability is stronger.This article takes the vision system of shopping mall service robots as the research object.Based on the investigation needs of Guangzhou Zhengjia Plaza and consulting literatur e in related fields,the following researches are mainly carried out:(1)A shopping mall service robot vision system integrating Faster R-CNN target detection model and generalized zero-sample learning model was established.First,through in-depth study of business requirements,design the overall framework of the target detection system that meets specific scenarios,and design the system hardware and software,including vision system carrier,camera selection,image acquisition and transmission,and software environment.Then,the robot target detection technology implementation route is condensed.The first step of the technical route uses the Faster R-CNN model to solve the detection tasks of multiple targets in the mall.The second step of the technical route uses the zero-sample learning model to achieve accurate pedestrian recognition.Finally,the integrated technology is tested and analyzed in the shopping mall environment to verify the effectiveness of the entire target detection system.(Chapter 2,Chapter 5)(2)The Faster R-CNN model is used to solve the detection tasks of multiple targets in shopping malls.First,the basic theoretical knowledge of Faster R-CNN target detection model is detailed.Then,the Faster R-CNN model is used to solve the detection tasks of multiple targets in shopping malls,and the two key steps of image preprocessing and image annotation in this scene are studied.Finally,the Tensor Flow deep learning framework is used to build the model and complete the training,and the Faster R-CNN target detection model that meets the requirements of shopping mall target detection is obtained.(Chapter 3)(3)A new generalized zero-shot pedestrian recognition model with knowledge transfer capability is proposed.First,the zero-sample image target recognition task for shopping malls was conceived and the corresponding pedestrian data set was produced.Then through the research on the three core technologies of zero sample learning,that is,image attribute feature extraction,semantic embedding space construction,and visual-semantic mapping,a new generalized zero shot learning model is constructed.Among them,for the problem of poor ma pping between CNN features and semantic space,the application of relational network connects CNN features and semantic embedding space together to make it better to achieve visual-semantic mapping.Finally,the use of accuracy thresholds allows the classi fier to select the best of the supervised deep learning model and the zero-sample learning model for target classification,so that the model has the role of knowledge transfer,improves the model's generalization ability,and validates the model used in t his paper The data set is 12.9% more accurate than the supervised deep learning model.(Chapter 4)...