Research And Implementation Of Human Behavior Recognition Technology Based On LiDAR

Human behavior recognition has a wide range of applications in the field of artificial intelligence and is gradually being applied in daily life.This thesis uses Li DAR as a hardware device for behavior recognition to collect point cloud images of human behavior.In this thesis,human behaviors are divided into two broad categories of gestures and actions,and the data are collected separately for behavior recognition.Although convolutional neural networks have disadvantages such as long training time and high hardware overhead,they have superior performance and are widely used in various fields such as speech recognition,image recognition,and target detection.In this thesis,we use convolutional neural networks to realize human behavior recognition based on Li DAR by analyzing and studying domestic and foreign technologies about behavior recognition.The main work of this thesis is as follows.To address the problems of high complexity and high training overhead of convolutional neural network models in Li DAR human pose recognition,this thesis proposes an improved Efficient Net-based human pose recognition model.Firstly,the model is briefly described,and secondly,the Efficient Net model architecture is analyzed,and the Efficient Net is improved in terms of the overall structure and activation function to improve the model recognition efficiency.The dataset used in this thesis is much smaller in scale compared with the Image Net dataset,and the Efficient Net model is too complex when the neural network is trained using the small-scale dataset;through extensive experiments,the number of major modules in this model is reduced and the Efficient Net is streamlined in terms of depth without decrease in the accuracy of the model.After the adjustment of Efficient Net,the number of network layers is reduced,and the Swish activation function no longer has advantages;this thesis uses the Leaky Re LU activation function instead of the Swish activation function,and the experiments prove that the Leaky Re LU activation function works better in the network with shallow depth.For the problem that the human action features are difficult to extract due to too many voids and noise in the 3000 ms point cloud map in Li DAR human action recognition,this thesis firstly selects suitable structural elements to fill the large area voids in the point cloud map by morphological closed operation;secondly uses iterative guided filtering algorithm to remove the noise and non-essential small structures in the point cloud map to retain the human action edge features while removing irrelevant and redundant information in the image;then manually extract the edge feature information of human action using the Canny edge detection algorithm,the removal of voids and noise provides the edge detection,otherwise a large number of voids and noise are mistaken for edge information,which affects the efficiency of feature extraction;finally combine the manual feature extraction with the automatic feature extraction in deep learning,using the processed single channel image for the improved Efficienct Net is trained,which effectively improves the accuracy of action recognition and saves some training time.A prototype LIDAR-based human behavior recognition system is designed and implemented on the basis of posture and motion recognition methods.The framework of the system and the specific process of human behavior recognition are described in detail,and the functions of different modules of the system are analyzed and explained in detail.Finally,the performance of the human behavior recognition system is tested,and the recognition results of the system are intuitively displayed through a simple and friendly visual operation interface to prove the feasibility of the human behavior recognition model and system in this thesis.