Using neural network classifiers for the implementation of the vehicle identification system
DOI:
https://doi.org/10.34185/1562-9945-2-133-2021-15Keywords:
Python, PytorchAbstract
The work is devoted to the study of neural network classifiers for the implementation of a vehicle identification system. The problem of recognition of light signals of vehicles was solved in the work.
The light version of YOLOv3 was used as a vehicle detector, and the MobileNetv2 ar-chitecture was adapted to the light signal classifier. The models were trained on several da-tasets, brought to a single format.
The resulting quality of the models is good enough to prove that the system works. Fur-ther quality improvement is possible by increasing the volume of the training sample and more accurate selection of model hyperparameters. From the point of view of the speed of work, the recognition system showed an acceptable result.
From the point of view of the speed of work, the recognition system showed an acceptable result. The entire process of processing a frame can be done in a reasonable amount of time on budget equipment. However, for a full-fledged embedded solution, it is necessary to further speed up the processing process, while reducing the consumption of graphics memory and the load on the processor.
When checking the operability on the recordings from the DVR, the system spoke well, but sometimes incorrect results were observed for distant cars.
References
NVIDIA DRIVE AGX / nvidia.com – URL: https://www.nvidia.com/enus/self-driving-cars/drive-platform/hardware/
Wei L., Hong B. Vision-Based Method for Forward Vehicle Brake Lights
Recognitio. // International Journal of Signal Processing. – 2015. – Vol.8. – No.6. – pp. 167-180.
Jian-Gang W., Lubing Z. Real-time Vehicle Signal Lights Recognition with HDR Camera. / Conference: 2016 IEEE International Conference on Internet of Things (iThings) and IEEE Green Computing and Communications (GreenCom) and IEEE Cyber, Physical and Social Computing (CPSCom) and IEEE Smart Data (SmartData) // ieeexplore.ieee.org. – URL:
https://ieeexplore.ieee.org/document/7917112
Yuki O., Yoshihiro S. HSV Color Space Based Lighting Detection for Brake Lamps of Daytime Vehicle Images. // Journal of Computers. – 2019. – Vol.14. – No.1. – pp. 25-30.
Wang Z., Huo W. Performance Evaluation of Region-Based Convolutional Neural Net-works Toward Improved Vehicle Taillight Detection. // Applied Sciences. – Vol.9. – No.18:3753. – 2019. – pp. 1-19.
Krizhevsky A., Sutskever I. ImageNet Classification with Deep. // Advances in neural in-formation processing systems. – 2012. – Vol.25. – No.2. – pp. 1-9.
Szegedy C., Liu W. Going deeper with convolutions. // 2015 IEEE Conference on Com-puter Vision and Pattern Recognition (CVPR). – 2015. – pp. 1-9.
Ren S., He K. Faster R-CNN: Towards Real-Time Object Detection with
Region Proposal Networks. // IEEE Transactions on Pattern Analysis and Machine
Intelligence. – 2017. – Vol.39. – Issue 6. – pp. 1137 - 1149.
Redmon J., Farhadi A. YOLOv3: An Incremental Improvement. – URL:
https://arxiv.org/abs/1804.02767
NVIDIA TensorRT / developer.nvidia.com –
URL: https://developer.nvidia.com/tensorrt
Open Neural Network Exchange / onnx.ai – URL: https://onnx.ai/
Bewley A., Ge Z. Simple Online and Realtime Tracking. – URL:
