Design And Research Of Pedestrian Detection System Based On Embedded Platform

As one of the most important service objects in the history of computer vision development and application,pedestrian detection technology is that the machine correctly distinguishes people from their environment through cameras or other sensors.In many application scenarios,pedestrian detection plays an important role in the interaction between machines and humans.For example,in the application areas of intelligent security,unmanned cars,and intelligent robots.In recent years,the object detection technology based on deep learning has developed rapidly.Compared with traditional pedestrian detection technology,the former has a higher detection accuracy and detection rate than the latter.In this paper,the pedestrian detection algorithm is improved based on the deep learning neural network model,and combined with the low cost,easy deployment and small size of the embedded platform,a pedestrian detection system based on the embedded platform is researched and implemented.In view of the specific implementation of the system,this article starts work from three aspects: research and improvement of pedestrian detection algorithms,system and hardware platform construction,design and implementation of applications and interfaces.Firstly,the lightweight target detection network MobileNet-SSD was selected as the basic network,and then research was conducted on the situation of the network’s poor ability to detect small targets,and improvements were made to the four aspects of the model: 1)A structural improvement The structure of sampled feature fusion makes up for the shortcomings of insufficient semantic features of the shallow network;2)replaces the network loss function to improve the problem of uneven proportion of positive and negative sample losses;3)improves the image augmentation algorithm and improves the small target object in the positive sample 4)Based on the statistical results of pedestrian images,the a priori frame parameters in the network are redesigned to improve the ability of the network a priori frame to match small target objects.Incremental experiments were performed on the four improved methods.The results show that compared with the previous model,the training loss and the detection ability of small target objects are significantly improved.Secondly,choose security monitoring as the application scenario of this experiment to design the overall framework of the system.The Linux operating system and Tengine,Qt and other third-party libraries are transplanted on the embedded platform,and a simple embedded pedestrian detection and monitoring platform is built as the experimental hardware platform in cooperation with the binocular camera and display.Finally,combined with the actual application requirements of the monitoring scenario,develop multi-threaded applications and graphical user interfaces on the embedded platform.Test the equipment in two ways.1)Overlay four improved methods one by one to test the accuracy of the model.Then by compressing the depth of the model appropriately,the average single frame detection time of the model on the embedded platform is increased to 240 ms.2)Perform actual testing in a monitoring scenario where only humans are detected and recorded.The results show that the storage space occupied by the system for one day is only one-twentieth of that of traditional monitoring,and the detection rate reaches 98%.This experiment achieves the expected results,and the basic functions of the system are stable and have certain practical application value.