A Research Of Embedded-based Pedestrian Detection,Tracking And Counting Systems

With the advancement and development of artificial intelligence technology,the performance of pedestrian-based detection,tracking and re-identification algorithms has been further improved,but the algorithms are still difficult to land.There are two main difficulties: first,most current deep learning neural network models only test on servers,and the models are difficult to run on embedded devices.Secondly,researchers have focused on a single problem,such as target detection and tracking,without considering the fusion between algorithms,which alone have significant limitations in real-life applications.In response to the above problems,this paper explores the application of neural network models to embedded devices.We select a lightweight model with excellent performance in pedestrian detection and tracking,and implement model pruning and quantification to further improve the algorithm's speed while ensuring accuracy.In addition,based on the urgent need for pedestrian information statistics in real life,a end-to-end system is designed in this paper,where users can obtain pedestrian counting information without any other operation.Finally,we adopt both short and long time pedestrian re-identification strategies to solve the ID switching problem that may occur during the pedestrian tracking process.We allow the system to automatically update the database ID of pedestrians to ensure the uniqueness of the pedestrians' ID in the database.We count pedestrians by the number of IDs they have and therefore also ensure the accuracy of the pedestrian counts.The results of the final experiment also show that the algorithm for deep learning is not impossible to run on embedded devices.We need to get the most out of our algorithms on embedded devices,combining the best of both worlds to get real utility value.