Traffic sign recognition is an important problem for autonomous cars and driver assistance systems. With recent developments in the field of machine learning, high performance can be achieved, but typically at a large computational cost. This thesis aims to investigate the relation between classification accuracy and computational complexity for the visual recognition problem of classifying traffic signs. In particular, the benefits of partitioning the classification problem into smaller sub-problems using prior knowledge in the form of shape or current region are investigated. In the experiments, the convolutional neural network (CNN) architecture MobileNetV2 is used, as it is specifically designed to be computationally efficient. To incorporate prior knowledge, separate CNNs are used for the different subsets generated when partitioning the dataset based on region or shape. The separate CNNs are trained from scratch or initialized by pre-training on the full dataset. The results support the intuitive idea that performance initially increases with network size and indicate a network size where the improvement stops. Including shape information using the two investigated methods does not result in a significant improvement. Including region information using pretrained separate classifiers results in a small improvement for small complexities, for one of the regions in the experiments. In the end, none of the investigated methods of including prior knowledge are considered to yield an improvement large enough to justify the added implementational complexity. However, some other methods are suggested, which would be interesting to study in future work.

The traffic sign detection and recognition system is an essential part of the driver warning and assistance systems. In this thesis, traffic sign recognition system is studied. We considered circular, triangular and square Turkish traffic signs. For detection stage, we have two different approaches. In first approach, we assume that the detected signs are available. In the second approach, the region of interest of the traffic sign image is given. Traffic sign is extracted from ROI by using a detection algorithm. In recognition stage, the ring-partitioned method is implemented. In this method, the traffic sign is divided into rings and the normalized fuzzy histogram is used as an image descriptor. The histograms of these rings are compared with the reference histograms. Ring-partitions provide robustness to rotation because the rotation does not change the histogram of the ring. This is very critical for circle signs because rotation is hard to detect in circle signs. To overcome illumination problem, specified gray scale image is used. To apply this method to triangle and square signs, the circumscribed circle of these shapes is extracted. Ring partitioned method is tested for the case where the detected signs are available and the region of interests of the traffic sign is given. The data sets contain about 500 static and video captured images and the images in the data set are taken in daytime.

No description available.

In recent years, AI has developed significantly and become more popular than ever before. The applications of AI are expanding, making knowledge about its application and the systems it can be applied to more important. This project compares and evaluates deep learning models for object detection of traffic signs on the embedded systems Nvidia Jetson Nano and Raspberry Pi 3 Model B. The project compares and evaluates the models YOLOv5, SSD Mobilenet V1, FOMO, and Efficientdet-lite0. The project evaluates the performance of these models on the aforementioned embedded systems, measuring metrics such as CPU usage, FPS and RAM. Deep learning models are resource-intensive, and embedded systems have limited resources. Embedded systems often have different types of processor architectures than regular computers, which means that some frameworks and libraries may not be compatible. The results show that the tested systems are capable of object detection but with varying performance. Jetson Nano performs at a level we consider sufficiently high for use in production depending on the specific requirements. Raspberry Pi 3 performs at a level that may not be acceptable for real-time recognition of traffic signs. We see the greatest potential for Efficientdet-lite0 and YOLOv5 in recognizing traffic signs. The distance at which the models detect signs seems to be important for how many signs they find. For this reason, SSD MobileNet V1 is not recommended without further trai-ning despite its superior speed. YOLOv5 stood out as the model that detected signs at the longest distance and made the most detections overall. When considering all the results, we believe that Efficientdet-lite0 is the model that performs the best. / Under de senaste åren har AI utvecklats mycket och blivit mer populärt än någonsin. Tillämpningsområdena för AI ökar och därmed blir kunskap om hur det kan tillämpas och på vilka system viktigare. I det här projektet jämförs och utvärderas djupinlärningsmodeller för objektdetektering av trafikskyltar på de inbyggda systemen Nvidia Jetson Nano och Raspberry Pi 3 Model B. Modellerna som jämförs och utvärderas är YOLOv5, SSD Mobilenet V1, FOMO och Efficientdet-lite0. För varje modell mäts blandannat CPU-användning, FPS och RAM. Modeller för djupinlärning är resurskrävande och inbyggda system har begränsat med resurser. Inbyggda system har ofta andra typer av processorarkitekturer än en vanlig dator vilket gör att olika ramverk och andra bibliotek inte är kompatibla. Resultaten visar att de testade systemen klarar av objektdetektering med varierande prestation. Jetson Nano presterar på en nivå vi anser vara tillräckligt hög för användning i produktion beroende på hur hårda krav som ställs. Raspberry Pi 3 presterar på en nivå som möjligtvis inte är acceptabel för igenkänning av trafikskyltar i realtid. Vi ser störst potential för Efficientdet-lite0 och YOLOv5 för igenkänning av trafikskyltar. Hur långt avstånd modellerna upptäcker skyltar på verkar vara viktigt för hur många skyltar de hittar. Av den anledningen är SSD MobileNet V1 inte att rekommendera utan vidare träning trots sin överlägsna hastighet. YOLOv5 utmärkte sig som den som upptäckte skyltar på längst avstånd och som gjorde flest upptäckter totalt. När alla resultat vägs in anser vi dock att Efficientdet-lite0 är den modell som presterar bäst.

In the area of Traffic Sign Recognition (TSR), deep learning models are trained to detect and classify images of traffic signs. The amount of data available to train these models is often limited, and collecting more data is time-consuming and expensive. A possible complement to traditional data acquisition, is to generate synthetic images with a generative machine learning model. This thesis investigates the use of denoising diffusion probabilistic models for generating synthetic data of one or multiple traffic sign classes, when providing different amount of real images for that class (classes). In the few-sample method, the number of images used was from 1 to 1000, and zero images were used in the zero-shot method. The results from the few-sample method show that combining synthetic images with real images when training a traffic sign classifier, increases the performance in 3 out of 6 investigated cases. The results indicate that the developed zero-shot method is useful if further refined, and potentially could enable generation of realistic images of signs not seen in the training data.