PROJECT TITLE :

Automatic Traffic Sign Detection and Recognition Using SegU-Net and a Modified Tversky Loss Function With L1- Constraint

ABSTRACT:

Autonomous vehicle technology relies heavily on traffic sign detection. Researchers have been inspired to employ neural networks to handle this task as a result of recent improvements in deep learning techniques. We examine traffic sign detection as an image segmentation problem in this paper and offer a deep convolutional neural network-based solution to tackle it.

To this purpose, we propose the SegU-Net, a new network that we create by combining the state-of-the-art segmentation architectures SegNet and U-Net to detect traffic indications in video sequences. Instead of the intersection over union loss that is generally used to train segmentation networks, we employ the Tversky loss function restricted by an L1 term to train the network. To identify the discovered indications, we employ a different network inspired by the VGG-16 architecture.

The networks are trained on the CURE-TSD dataset's challenge-free sequences. Our suggested network exceeds state-of-the-art object detection networks by a considerable margin, with precision and recall of 94.60 percent and 80.21 percent, respectively, which is the current state of the art on this section of the dataset.

The network is also put to the test on the German Traffic Sign Detection Benchmark (GTSDB) dataset, where it achieves precision and recall of 95.29 and 89.01 percent, respectively. This is comparable to the performance of the object detection networks stated above. These findings demonstrate the architecture's generalizability and applicability for robust traffic sign detection in autonomous cars.