Redes neurais convolucionais para a detecção de objetos
Carregando...
Data
Autores
Título da Revista
ISSN da Revista
Título de Volume
Editor
Pontifícia Universidade Católica de Campinas (PUC-Campinas)
Resumo
Carros autônomos (ACs) e sistemas avançados de assistência ao motorista (ADAS) contam com
redes neurais convolucionais (CNNs) para detecção de objetos. No entanto, a degradação da
imagem causada por condições climáticas adversas, como chuva, neve e neblina, pode diminuir
o desempenho de uma CNN. Assim, este trabalho apresenta o desenvolvimento de uma técnica
de processamento de imagem com o objetivo de mitigar tal problema. Primeiramente, após uma
extensa avaliação de modelos para detecção de objetos, nossa escolha recaiu sobre a YOLOv3,
devido a seu compromisso entre precisão e tempo de inferência. Posteriormente, o treinamento
e teste de uma CNN YOLOv3 foi investigado para carros, semáforos, semáforos, pedestres e
ciclistas/motociclistas. O desempenho foi avaliado estimando-se a precisão média e média da
precisão média (mAP) para cada uma das classes de objetos mencionadas. Foi implementada
uma técnica de pré-processamento baseada em OpenCV para mitigar a degradação imposta por
condições climáticas adversas. Em vista disso, os filtros do OpenCV de erosão, dilatação e joint
bilateral filter foram considerados durante o treinamento e testes dos conjuntos de dados
Berkeley DeepDrive (BDD100K) e Detection in Adverse Weather Nature (DAWN). O trabalho
desenvolvido apresenta os benefícios potenciais do uso de filtros OpenCV como aumento de
dados durante treinamento e testes. Nossos resultados mostram uma melhora em torno de 3% no
mAP durante os testes com DAWN.
Autonomous cars (ACs) and advanced driver-assistance systems (ADAS) have relied on convolutional neural networks (CNNs) for object detection. However, image degradation caused by adverse weather conditions like rain, snow, and fog can decrease the performance of a CNN. So, this paper presents the development of an image-processing technique aimed to mitigate such a problem. First, after an extensive evaluation of models for object detection, our choice fell on YOLOv3, because of its compromise between precision and inference time. Afterwards, the training and test of a YOLOv3 CNN was investigated for cars, traffic signals, traffic lights, pedestrians, and riders. Performance was evaluated by estimating the average and mean average precision (mAP) for every one of the mentioned object classes. An OpenCV based pre-processing technique to mitigate the degradation imposed by adverse weather conditions was implemented. Hence, the OpenCV filters of erosion, dilation and joint bilateral filter were considered during training and tests of the datasets Berkeley DeepDrive (BDD100K) and Detection in Adverse Weather Nature (DAWN). The developed work presents the potential benefits of OpenCV filters use as data augmentation during training and testes. Our results show an improvement around 3% in mAP during tests with DAWN.
Autonomous cars (ACs) and advanced driver-assistance systems (ADAS) have relied on convolutional neural networks (CNNs) for object detection. However, image degradation caused by adverse weather conditions like rain, snow, and fog can decrease the performance of a CNN. So, this paper presents the development of an image-processing technique aimed to mitigate such a problem. First, after an extensive evaluation of models for object detection, our choice fell on YOLOv3, because of its compromise between precision and inference time. Afterwards, the training and test of a YOLOv3 CNN was investigated for cars, traffic signals, traffic lights, pedestrians, and riders. Performance was evaluated by estimating the average and mean average precision (mAP) for every one of the mentioned object classes. An OpenCV based pre-processing technique to mitigate the degradation imposed by adverse weather conditions was implemented. Hence, the OpenCV filters of erosion, dilation and joint bilateral filter were considered during training and tests of the datasets Berkeley DeepDrive (BDD100K) and Detection in Adverse Weather Nature (DAWN). The developed work presents the potential benefits of OpenCV filters use as data augmentation during training and testes. Our results show an improvement around 3% in mAP during tests with DAWN.