Research On Detection Method Of Water Body Area In Road Scene Image Based On Deep Learning

Unmanned Ground Vehicles(UGV)have become a hot research topic in recent years.For UGV,environment perception is the key of reliable operation.Unlike Unmanned Aerial Vehicles(UAV),UGV usually faces more complex terrain situations,unstable obstacle types and uncertain factors.In the open environment,accidentally driving into puddles,mires and other terrain may result in UGV to break down directly,causing immeasurable losses.Therefore,the accurate detection of water puddle on roads has important practical significance.Based on the characteristics of the water puddle detection problem in the road scene,we fully utilize the physical characteristics of the water puddle and the strong capability of deep learning,and the following three aspects are investigated in this paper:(1)A water puddle detection model of road images which named URA-Net that combines reflection attention unit(RAU)and self-attention mechanism was proposed.By adding residual convolution blocks and up-sample convolution blocks on basis of U-Net,the model has achieved performance beyond the existing deep neural networks for water puddle detection.At the same time,we combine reflection attention unit and self-attention mechanism to further boost the performance.The F1-Measure of ‘Both Road' in ‘Puddle-1000' reaches 87.18%.The experiment results show that the dual attention module can better capture feature dependency and improve the representation ability of water semantic features comparing to the single use of RAU.(2)A water puddle detection model for road scene images which named RWD-GAN based on two-generators adversarial learning was proposed.RWD-GAN has a novel structure of dual generators and one discriminator.The dual generators still use a model architecture similar to URA-Net,but they are modified for specific goals.It is hoped that through the adversarial learning between the generators and the discriminator as well as the two generators,the information transmission will be promoted and achieve a better balance between two competing goals.Meanwhile,we give a kind of discriminator design suitable for water puddle detection of road scene images.The F1-Measure of ‘Both Road' in ‘Puddle-1000' reaches 88.54%,which shows that the dual generators adversarial learning can better improve the performance of water puddle detection.(3)A kind of water puddle detection method based on domain adaptation which named Seg-DANet was proposed.Seg-DANet is composed of an encoder,a decoder,and two discriminators.The encoder is responsible for giving the potential representation of the water puddle features,the decoder is responsible for combining the domain code to generate the corresponding segmentation results,and the two discriminators act on the source domain and the target domain respectively.The combination of the encoder,decoder and discriminators realize the extraction of domain-invariant feature representations while boosting image segmentation tasks.Compared with directly using the model trained on the source domain,the performance of unsupervised domain adaptation is improved by 10.5%,and the performance of semi-supervised domain adaptation is improved by 36.8%,which proves that domain adaptation can effectively enhance the generalization ability and environmental adaptability of model.