Design And Research Of A Traffic Flow Detection System Based On SSD＿MobileNet Algorithm

Traffic flow detection technology can provide accurate information on road conditions in real time to help drivers plan their routes and improve road utilisation.Traditional traffic flow detection methods are unable to meet the changing needs.Video detection technology has slowly started to develop with the rise of machine learning.In this paper,traffic flow detection is studied,using RK3399 Pro as the detection platform,porting SSD＿MobileNet vehicle detection algorithm and improving SSD＿MobileNet vehicle detection algorithm and combining with DeepSort algorithm for vehicle tracking and counting.Finally,road tests were conducted.In order to achieve the above objectives,this paper focuses on the following aspects of research.Hardware system platform design.Mainly including Boot Loader,system files and dependency environment construction and porting,serial debugging and model conversion tool RKNN-Toolkit installation and configuration.Learning deep learning target detection and target tracking algorithms,for the singlestage target detection algorithm SSD＿Mobilenet in the training process led to the imbalance between the front and back background samples,the Focal Loss function is introduced into the confidence loss function,adjusting the value of different weights to target difficult samples,to optimise the network learning ability.At the same time,regression prediction is performed separately between the feature layers of the network,and FPN is introduced for the fusion of high-level and bottom-level feature information to give an improved algorithm network structure and enhance the detection capability of the algorithm for small targets.Training of improved algorithms and testing of algorithm accuracy and speed.Combine vehicle detection module,tracking and counting module to achieve traffic counting.Using the model conversion tool developed by Rexchip,the trained algorithm model was converted into an RKNN file and ported to the ARM.Assigned for execution on the NPU.Testing of a traffic detection system ported with the improved SSD＿MobileNet algorithm and the DeepSort tracking algorithm consisted of importing recorded video into memory in advance and taking video in real time using a camera and analysing the results.It has been proven that this system can detect and track vehicles more accurately,with an accuracy rate of 92.76%.It provides a new idea for the development of traffic flow detection system.