Real-time attribute based deep learning network for traffic sign detection

• Main Authors: Mohamed Elhawary, Hossamelden, Suddamalla, Upendra, Shapiai, Mohd. Ibrahim, Wong, Anthony, Zamzuri, Hairi
• Format: Conference or Workshop Item
• Published: 2022
• Subjects: T Technology (General)
• Online Access: http://eprints.utm.my/id/eprint/98694/; http://dx.doi.org/10.1109/ICITEE56407.2022.9954110

Traffic sign detection is one of the key components of the Advanced Driving Assistant System (ADAS), which aims to detect and classify street signs in real time. However, traffic sign detection has challenges in real applications requiring high precision and real-time recall. Those challenges are due to the small object size and class imbalance. Recently, researchers have proposed several techniques to improve the detection quality by enriching the features through a multiscale network, introducing attention mechanisms and augmentation techniques to improve the features of tiny objects. To overcome class imbalance researchers proposed cascaded networks and various loss functions. However, those existing techniques and mechanisms added more complexity to the model. Meanwhile, the imbalance affects single-stage networks such as YOLO, which causes a lower recall for minor classes. We proposed a new training method for a single-stage detection network, known as Real Time Attribute Based Deep Learning Detection Network (Real Time-Attribute DL). We introduced new attributes to the loss and Non-Maximum Suppression (NMS) to reduce the class number by categorizing it based on the shape of the traffic sign while maintaining the same number of classes. Our proposed method extends the YOLO detection head to have four main parameters: objectiveness, regression, class, and attribute. We modify the loss function to train the network jointly between class and attribute. We validate our proposed technique with Tsinghua-Tencent 100K(TT100K) as a benchmark dataset. The results show that our proposed technique improves the recall index from 85.85% to 94.26% in yolov4-tiny-31 with a 0.8% improvement in precision and improves the recall index from 93.51% to 96.68% in yolov4 with a drop by 2% in precision without adding extra complexity to the main network. The proposed technique offers a better recall index than the baseline, especially for imbalanced datasets such as TT100K datasets.