Lightweight Target Recognition Technology Based On Multimodal Fusion

With the rapid development of science and technology,target recognition technology is widely used in many fields,such as medical treatment,agriculture,autonomous driving,security monitoring,environmental monitoring,etc.,which makes human life safer and more convenient.In the field of security and military,target recognition in the field environment is still challenging,mainly reflected in the following aspects:First,the situation in the field environment is complex and changeable.Currently,most target recognition technologies are mostly based on a single means,and target recognition is susceptible to interference,with ambiguity and uncertainty.Secondly,most target recognition relies on complex feature extraction processes and a large number of floating-point calculations.It is difficult for small and micro sensors to have high computational power.In outdoor environments,it is necessary to ensure the accuracy of target recognition,while reducing computational complexity and saving computational costs.In addition,in the face of unexpected events such as target intrusion,security personnel can make decisions more efficiently by knowing the on-site situation as soon as possible,and the real-time performance of the system is also a top priority.Aiming at the above problems,this paper studies the lightweight target recognition technology based on multimodal fusion in the field environment,and designs and implements the intelligent monitoring of lightweight targets using multimodal sensor collaboration.First of all,in view of the complex situation in the field and the vulnerability of single means to interference,this paper studied three methods of target recognition,and studied and implemented feature extraction and target recognition based on YOLOv5s image mode,DBSCAN algorithm based millimeter wave mode,and MFCC based sound mode.Secondly,in order to give better play to the advantages of each mode,after obtaining the feature information of a single mode,the SENet based multimodal fusion method is used to closely fuse the three modal features,enhance the obvious features,weaken the indistinct features,and achieve a more accurate and robust target recognition system.Finally,in order to ensure real-time decision-making and reduce computational complexity,the Ghost module is used to perform lightweight model compression on the convolutional neural network,reducing a large number of parameters while improving computational speed.To facilitate the use of security personnel,based on the above algorithms,this article implements a target identification module and a device management module,completes real-time visual viewing of target identification results and raw data,viewing and managing sensor information,and provides a visual platform for environmental monitoring.