Safety Monitoring And Early Warning System For Crowded Places Neural Based On Computation Stick (NCS2)

With the development of society,people flow monitoring plays a more and more important role in safety prevention,urban construction,market decision-making and other aspects,especially in the crowded places such as squares,stations,department stores,etc.,where people are too concentrated,it is easy to cause safety accidents.If the traditional method is used for manual monitoring,in the face of a large number of large surveillance videos,long-term observation will lead to administrator fatigue and huge energy consumption.It is also possible that management personnel leave the monitoring due to various emergencies,resulting in omissions.Therefore,the crowd monitoring and early warning system of stadiums and gymnasiums is particularly important.In order to better ensure the safety of crowds in dense places,this topic adds a neural computing stick to the Raspberry Pi 4B equipped with Raspberry Pi OS system,uses the deep learning technology to detect the target and calculate the density of people,obtains the density of each part of the crowd,displays the crowd density through the thermal map,and voice prompts the evacuation direction in the crowded area.The research of this system is carried out from the following aspects:Firstly,the system is analyzed macroscopically,and its characteristics and requirements are listed.This paper introduces deep learning and edge computing,and concludes that edge computing has many advantages.After comparing Raspberry Pi 4B with NVIDIA Jetson Nano,it selects Raspberry Pi 4B and gives the reasons,and makes a comprehensive analysis on its performance and characteristics;The Raspberry Pi OS operating system and neural computing stick used in this design are also introduced;The advantages and disadvantages of different reasoning frameworks are compared,and the best reasoning framework Open VINO is selected.Secondly,the construction method of pedestrian safety monitoring model is introduced.After searching the public data set,the data set is preliminarily processed,the data set is divided,and the model and parameters are trained on the PC side.This paper introduces three target detection algorithms,MSCNN,YOLOv5 and YOLOv6,and finally selects YOLOv5 as the algorithm of this subject,and improves its frame rate through inference framework and neural computing stick.Then it introduces the implementation method of safety monitoring and early warning system in crowded places.The camera with a FOV of 60 degrees is selected to train the information acquisition module and obtain test samples.This paper improves the target detection algorithm YOLOv5 based on crowd counting algorithm,and uses this algorithm to train and reason the safety monitoring model of crowded places.And the display and early warning module is realized.When the system detects that the crowd has reached the early warning threshold,the screen will display the early warning information and the corresponding treatment suggestions,so that the staff can take action in time.Finally,the relevant test work is carried out on the passenger flow safety monitoring system of YOLOv5,which is based on the detection of dense crowd counting,and the functions of each module are verified.The power consumption of Raspberry Pi 4B and PC are tested respectively,and the results are compared.The CPU usage,accuracy,and frame rate were compared,and the test results were analyzed.The target detection model selected after the comparison was imported into Raspberry Pi 4B for comparative experiments.Direct detection,Open VINO plus 1 NCS2,and Open VINO plus 2 NCS2 were compared.The system uses a camera to read the crowd image and display the density thermogram.When the number of people is more than 10,the screen displays "Warning" and prompts evacuation through the voice module.The experimental results show that the use of Open VINO and neural computing sticks can improve the reasoning speed of YOOv5,and the frame rate of direct detection is only 1.549 fps.The frame rate after adding one neural computing stick is126.9% higher than that of direct detection,and the frame rate after adding two neural computing sticks is 286.5% higher than that of direct detection.