Research On Target Detection And Prediction Of Urban Traffic

Urban traffic object detection and short-term traffic flow prediction are important components of intelligent connected transportation systems and have become a research hotspot in the field of road network traffic monitoring.However,in practical application scenarios,traditional object detection algorithms are difficult to directly deploy in the road network due to limited computational and storage resources of road monitoring devices.Meanwhile,urban daily traffic flow exhibits nonlinear,non-stationary,and time-dependent characteristics,leading to significant errors in traffic flow prediction.To address these issues,this study focuses on research in the following areas: lightweight design of object detection models and combined modeling of shortterm traffic flow prediction.The main research contents are as follows:(1)Research has been conducted around intelligent connected transportation systems.Due to the limited computational and storage resources of onboard units and roadside monitoring devices,directly deploying the traditional YOLOv4 model poses challenges in ensuring the efficient and adaptable of urban traffic object detection.To address this,a lightweight model for urban traffic object detection is proposed.The model begins by reconstructing the YOLOv4 backbone network based on the Ghost Net lightweight network to reduce model calculation.Additionally,The threedimensional Sim AM attention mechanism is integrated into the Ghost Bottleneck(GBN)component to enhance the model’s high-resolution capability for critical regions.Furthermore,the YOLOv4 feature fusion network is improved using deep separable convolutions to further reduce model parameters.Lastly,the label smoothing algorithm is employed to decrease the weight of true sample labels in the loss function calculation,thereby improving image classification accuracy.Experimental results demonstrate that the Edge-YOLOv4 model reduces parameters by approximately 82% compared to YOLOv4,increases detection speed by 7.06 frames/s,and verifies the robustness of target detection across different scenarios.(2)Due to the nonlinear,non-stationary,and time-dependent characteristics of urban traffic flow data,traditional algorithms often result in significant errors in traffic flow prediction.In this regard,a combined prediction model for short-term traffic flow on urban highways is proposed.The model first applies the improved fully adaptive noise-ensemble empirical mode decomposition(ICEEMDAN)method to decompose the original traffic flow sequences,refining the non-stationary characteristics of the traffic flow series.Next,bidirectional gated recurrent units(Bi GRU)are utilized to capture the temporal correlations within the traffic flow sequences.Lastly,an improved sparrow search algorithm based on dynamic adaptive distribution mutation is employed to iteratively optimize the weight parameters of the Bi GRU network,avoiding local optima in shortterm prediction results.Experimental results demonstrate that the improved traffic flow prediction model achieves an average absolute error of 10.98,average absolute percentage error of 10.12%,and root mean square error of 12.42.The model also exhibits good generalization performance across different datasets.(3)To address the insufficient application of traffic object detection and flow prediction in practical scenarios,as well as the problem of abstract traffic flow data being difficult for travelers to understand and utilize,a traffic visualization system specifically designed for urban transportation has been developed and implemented.This system comprises core modules such as system management,data collection,object detection,and traffic flow prediction.Additionally,the functional module design includes features like area heatmaps,traffic congestion rankings,traffic flow query,and prediction.This ensures the diverse needs of users and provides a data foundation and platform support to solve urban traffic congestion and facilitate user travel.