Research On Semantic Segmentation Method For Road Scene Understanding

Semantic segmentation is one of the key technologies for scene understanding.It has a wide range of application scenarios in many fields such as image retrieval,driver assistance,and intelligent surveillance.Road scene understanding applications,such as driver assistance system,impose higher requirements on semantic segmentation,which requires a semantic segmentation method with faster segmentation,lower resource overhead,and more accurate segmentation results.Recently,the deep learning method has achieved a series of outstanding achievement on the semantic segmentation problem.But,on the other hand,it leads to a more and more complex structure of the model.These kind of semantic segmentation methods could not work well in a road scene understanding application.In addition,the training of the network often requires a large number of training data.It means a huge work of pixel-wise annotation.The objective of paper is mainly to implement an efficient semantic segmentation network model and design a high-precision automatic tagging method to solve aforementioned problems.Main research and achievements in this paper are concluded as follow:A simplified convolutional neural network for semantic segmentation is implemented in this paper which has higher learning efficiency,better segmentation accuracy and lower memory overhead.This model,which is inspired from VGG-16 network and Seg Net,have been improved with a better structure and units.It uses PRe LU as its activation function which can help to get better learning efficiency.And this model introduces Dropout layers into network structure to reduce overfitting.With these improvements,the model achieves higher segmentation accuracy and training efficiency as a result.An automatic annotation method for semantic segmentation training data based on super-pixel to solve the problem on manual annotation is proposed in this paper.This method uses a novel feature designed in this paper to perform super-pixel classification according to color,texture,gradient and location of super-pixels divided by SLIC method.This method annotates all the pixels with the same label as the super-pixel they containedin.This method eliminates the complicated training process and provides a large number of supervised training data for automatic annotation through superpixel segmentation.It can reduce the requirement for the number of ground truth while guaranteeing the annotation accuracy.As a result,this method can help to reduce the overhead of manual annotation greatly.The experiments are performed on Cam Vid road scene dataset in this paper.The performance of designed model and method are proved to be good through the comparison with other existing methods.Results show that the semantic segmentation network model implemented in this paper has a significant improvement in segmentation accuracy and training efficiency.The improvement of the network structure is proved effective by further comparing the impact of different units on network performance.The superpixel-based automatic annotation method proposed in this paper also shows a good annotation result on Cam Vid dataset.Moreover,it can still maintain a high accuracy with less training data.