A Resource Constrained Pipeline Approach to Embed Convolutional Neural Models (CNNs)

• Ano de defesa: 2022
• Autor(a) principal: Silva, Rafael Andrade da
• Orientador(a): Prado, Bruno Otávio Piedade
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Programa de Pós-Graduação: Pós-Graduação em Ciência da Computação
• Palavras-chave em Inglês: Convolutional Neural Network (CNN); Edge devices; Quantization traffic signs; Knowledge distillation; Pruning; Traffic Sign Detection and Recogntion System (TSDR); German Traffic Sign Recognition Benchmark (GTSRB)
• Área do conhecimento CNPq: CIENCIAS EXATAS E DA TERRA::CIENCIA DA COMPUTACAO
• Link de acesso: https://ri.ufs.br/jspui/handle/riufs/19523

The objective of autonomous driving edge computer systems is to ensure the safety of Autonomous Vehicles (AV). However, this is extremely difficult. Advanced Driver Assistance Systems (ADAS) are of great importance in AV systems, as they increase the level of safety in vehicles. As vehicles become more connected, some ADAS features can be improved with the cooperation of the surrounding vehicles. For example, cooperative adaptive cruise control or a lane departure warning for all vehicles in the vicinity. Traffic Signal Detection and Recognition (TSDR) is a recent technology applied to intelligent driving responsible for identifying and recognizing traffic signs in the images captured by the vehicle’s sensors. TSDR systems have a wide range of applications. However, many of the proposed techniques use solutions based on expensive devices and are unsuitable for large-scale and low-cost edge computing solutions. Implementing these systems on OEM embedded platforms will provide the opportunity to create genuinely cost-effective and low-energy systems. In order to contribute to this research area, our study proposes not only the development of a convolutional neural network capable of performing the classification of vertical traffic signals but also the creation of a neural model compression pipeline. Based on the literature and experiments located through a systematic review, we chose to use the GTSRB dataset to evaluate the work. The pipeline has three stages: knowledge distillation, pruning, and quantization of neural models. The goal is to reduce the complexity of the final neural network, thus allowing the model to be embedded in a device with limited computational resources. The final models are evaluated considering performance metrics such as accuracy, precision, recall, F1-Score, inference time, and model size in bytes. Using the proposed methodology, our compressed CNN model achieved an accuracy of 85.91% and an F1-Score of 85.80%. The final model size was only 59 KB and the inference of a color image with a resolution of 32x32 pixels took only 80 ms to run in ESP32 and 83 ms to run in ESP32-S2, demonstrating the capability of this resource-constrained device to detect an image with a reasonable accuracy rate.