A Study On Traffic Event Detection Algorithm Based On Multi-Target Detection And Tracking

To achieve intelligent transportation and reduce the harm caused by traffic problems,accurate,reliable,and real-time traffic event information is urgently needed.Therefore,how to quickly and reliably detect and identify traffic events is one of the key technologies for achieving intelligent transportation.This paper focuses on traffic event detection technology and proposes a traffic event detection algorithm for multi-object detection and tracking.The research includes multiple aspects such as data acquisition,model optimization,and tracking framework to achieve accurate identification of traffic events.Finally,the designed algorithm is used to detect three types of traffic violations: pedestrian violations,vehicle lane violations,and vehicle wrong-way driving,to verify the effectiveness of the algorithm.The main work of this paper is as follows:(1)A data collection system was constructed to collect real-world traffic data in multiple scenarios.The data was cut into frames,screened,and key frames were extracted,and 10,000 image data frames were labeled.The label information was then statistically analyzed and formatted to construct a traffic participant detection dataset.A traffic event detection dataset was also constructed by combining network videos.(2)The YOLOX algorithm’s main network was designed for lightweight Ghost module introduction,reducing model parameters by 34.4% and computational complexity by 26.1%.Attention mechanisms were added between the main network and feature fusion network to enhance network feature extraction capabilities.The SIo U loss function was used to speed up network fitting speed,and the EFL loss function was used to increase the model’s focus on sparse categories.The model was then trained and verified using a self-made dataset.Experimental results show that the improved YOLOX algorithm achieved a 3.4%accuracy increase compared to the original YOLOX algorithm,with m AP reaching 93.5% and detection speed increasing from 47 FPS to 62 FPS.(3)The multi-object tracking algorithm Deep Sort was implemented,and an appearance feature extraction network was trained using the Re ID-type veri-wild dataset(15000 multi-angle images of vehicles)to extract vehicle appearance features,which were subsequently matched and tracked with the Kalman filter and data association matching algorithm.The traffic participant detection model and tracking model were combined and standardized in terms of data and image formats.The traffic participant tracking framework was optimized,a circular queue was proposed to store historical trajectory information,and the YOLOX+Deep Sort detection and tracking model was constructed.(4)Rules for pedestrian violations,vehicle lane violations,and vehicle wrong-way driving detection were designed and experimentally verified on a self-made dataset.The event detection rate for all three violations reached over 95%,proving that the traffic event detection algorithm designed in this paper can quickly and accurately identify traffic events.In summary,this paper proposes a new traffic event recognition algorithm for multi-object detection and tracking,including data acquisition,model optimization,and tracking framework,etc.The algorithm can improve the speed and stability of detection and tracking while ensuring accuracy,providing powerful support for the development of intelligent transportation systems.