Traffic sign recognition (TSR) has become an essential part of self-driving systems with computer vision algorithms recently. Several state-of-the-art computer vision techniques with the convolutional neural network (CNN) for TSR are designed in terms of signicantly managing the classication and detection accuracy. The research eld of TSR is strictly divided into traffic sign classication (TSC) and traffic sign detection (TSD).
In terms of improving hardware specication, the computer vision approaches with the CNN have demonstrated superior performance than the traditional methods. However, most approaches do not take into account some cases such as blur, illumination, or occlusion that can be seen in real applications. Therefore, the proposed method is designed to enhance classication performance by extracting the import regions of the images and minimizing its image loss under such cases. To achieve successful classication, a novel classication method with attention mechanism and convolutional-pooling are described with the augmented dataset.
The traditional CNN methods are designed with fully connected layer as a classier. However, fully connected layer draws an issue that spatial information from the feature maps is not preserved during the training process of the network and it results in performance degradation. An advanced CNN scheme is designed to perform robust classication with attentional-deconvolution modules (ADMs) and fully convolutional network for considering noisy environments.
The process of TSD is to predict candidate regions of the target objects by drawing multiple boxes in the images. Once, the bounding box for the target object is decided, the bounding box needs to be labeled by the following predened ground truth where TSC is applied. This dissertation proposes two classication approaches for traffic signs and expands them to actual traffic sign detection. A novel detection approach is specially designed to detect traffic signs considering its characteristics.
This dissertation proposes novel techniques for TSR based on the CNNs. The proposed CNNs are robust to noises and practical that do not require high-end specication of hardware. With previously studied CNNs, a new technology for TSD is introduced that is able to predict the locations of the traffic signs precisely without any complex strategy. Two different German traffic sign benchmarks are selected to provide the advanced performances of the proposed strategies. The dissertation contributes to practical insights of designing CNNs for the TSR, and these techniques are demanded of self-driving vehicle in the future.