Imitating Bat Three-dimensional Space Target Positioning System By Mobile Robot

With the development of artificial intelligence technology,sonar perception provides more and more perceptual information for artificial intelligence applications such as autonomous driving and intelligent robots.The use of ultrasound to locate targets in three-dimensional space is an important research direction of sonar perception systems,Its importance has become increasingly prominent.Studies have shown that bats can use their own special sound structure and auricle structure to form an ultrasonic target positioning system,which far exceeds the current artificial sonar system in terms of positioning accuracy and positioning sensitivity.But today,there are still few researches on bat-ear-like target positioning,especially the research on the recognition of target direction angle(pitch angle and direction angle).This paper mainly studies the target direction and angle recognition based on bat ears.On the basis of previous research,further research and improvement of the target direction and angle recognition of the imitation bat,combined with ultrasonic ranging technology,realized a mobile robot-based three-dimensional target positioning system,which can recognize the distance and pitch of the target Angle and azimuth.The specific research content of this paper is as follows:Firstly,an automatic data collection platform based on mobile robots was built.In terms of hardware system,the system is mainly composed of data acquisition module,mobile robot module and lift control module.The data acquisition module is composed of an ultrasonic horn and a bat-ear receiver,which is responsible for actively transmitting ultrasonic waves and collecting echo signals;the mobile robot module is composed of a four-wheel rotating car,lidar,industrial computer,etc.,which can automatically navigate to the specified location set in advance,And adjust its own posture;the lifting control module is mainly composed of a microcontroller,a motor and a set of pulleys,and automatically adjusts the height of the target according to the received instruction message.In terms of software system,the interaction between modules adopts TCP/IP protocol for communication,the application layer follows the websocket protocol,and programs are written to realize the communication and control of the automatic data collection platform.Based on this platform,we can collect large-scale data and provide data preparation for the recognition of the target direction and angle based on deep learning.Secondly,through experimental research,a deep learning model that fits the data in this paper is obtained.According to the characteristics of the data in this article,feedforward neural network(FNN),convolutional neural network(CNN)and recurrent neural network(RNN)are selected as learning models.For the feedforward neural network,the best structure of the model is obtained through experiments,and the influence of the momentum optimization algorithm and the adaptive learning rate optimization algorithm on the learning model is analyzed and compared;for the convolutional neural network,the characteristics of the traditional spectrogram On the basis,combined with the actual situation of this paper,creatively proposed the characteristics of multi-signal fusion spectrogram,and compared with the traditional splicing spectrogram characteristics;Long and short-term memory(LSTM)neural network and gated recurrent unit(GRU)neural network for experimental comparison and analysis.Through the research on the recognition model,the accuracy of target pitch angle recognition has been greatly improved.Finally,based on the above-mentioned data collection platform and the learning model obtained through training,a bat-like three-dimensional target positioning system based on a mobile robot was built.This paper uses the positioning system to collect data,and uses FNN,CNN and GRU as recognition models,respectively,to test the generalization ability of the above three models.The over-fitting mitigation method is used to fine-tune the above model;and a burst recognition method is proposed to improve the accuracy of model recognition.With reference to the recognition model of the pitch angle,the deep neural network recognition model of the target azimuth angle is trained to complete the recognition of the target direction angle.In the end,the accuracy of target direction angle recognition reached 82.25%(based on short burst).