| Abstrakt: |
This paper focuses on the smart security aspect of smart city initiatives, and specifically on road traffic monitoring. We describe the design and deployment of smart traffic monitoring in a pilot project located in the Selangor Cyber Valley, Selangor, Malaysia. Live roadside closed-circuit television footages are streamed to a remote network video recorder (NVR) and video management system. Videos from the NVR are also transmitted to a workstation at the Centre of Excellence for Advanced Sensor Technology for training, analysis, and inference with an artificial intelligence model. The proposed architecture allows for model updates without disruption to the running of the system. YOLOv7 is used for the object detection task of vehicle type identification, with enhancements for vehicle counting and brand/manufacturer classification. Identifying vehicles is crucial in regard to managing traffic flow, scheduling the movements of heavy vehicles, and identifying infrastructure improvements to alleviate congestion. The proposed model achieves precision, recall, and mean average precision (mAP) values of 80.6%, 88.3%, and 87.8%, respectively, surpassing the YOLOv7 model with h5 configuration by 3.6%, 9.3%, and 6.8%. The proposed model accurately detects 11 classes with mAP values of above 80%, except for buses, where the mAP is 70.6%. The similar characteristics of buses and vans (which are small objects, as they are captured from long distances) contribute to this lower accuracy, suggesting the need for more images and augmentation techniques to improve detection. The challenges encountered during and after deployment are addressed, and insights and recommendations are presented for future implementations. [ABSTRACT FROM AUTHOR] |
Full text from SpringerLink