Design of Automatic Truck Axle Counter Using Deep Learning on NVIDIA Jetson Nano

Abstract

To improve the quality of transportation, it is necessary to make right decisions based on accurate data. The flow of vehicles is important in the planning and operation of new roads and the modification of existing roads is needed to meet changes in traffic conditions. To get information about traffic characteristics, various information about traffic infrastructure and driver behavior is needed. The information obtained is analyzed to obtain traffic workload results. If the workload results are below the minimum service standard, a geometric change or arrangement of road space usage is required. The calculation of the number and classification of vehicles has been done manually at certain times. This requires a lot of human resources, so it is less efficient. Vehicle classification is based on the number of axles of the vehicle. This also applies to the payment system on toll roads. Therefore, this research created a real-time truck axle count system using deep learning on the NVIDIA Jetson Nano device. The system must be able to process in real-time with the use of compact hardware and use standard CCTV cameras. Based on the experimental results on 741 test image datasets, the average accuracy is 95.84% with a precision of 91.25%, recall is 79.13% and the processing speed reaches approximately 25 fps on live network camera.