Outdoor Scene Semantic Segmentation Algorithm Based On Convolutional Neural Networks For Smart Car

In recent years,smart driving has become increasingly hot.The related technology of intelligent driving has been gradually shifted from the development stage to market applications.Semantic segmentation technology can provide rich outdoor scene information for smart cars,it improve the cognitive ability of intelligent vehicles,it provides reliable technical support for decision control of smart cars,and the algorithm is robust,so the scene semantic segmentation is at the core in the smart car technology,and it has a wide range of application value.But because of the problem of the low accuracy and slow speed of the traditional semantic segmentation model,making it difficult to actually use.In response to this problem,this paper focuses on the method to improve the precision and speed of the algorithm.According to this theory,Based on the convolution neural network we have designed the FusNet and DeepSemNet two models.The main contributions of this paper are as follows.This paper discusses the input image size,incentive function and pooling,Which have the greatest impact on the speed and accuracy of the network.In this paper,we set the classic segmentation model as experimental model,and these models were trained and tested on the CamVid and ADE20 K databases,respectively.By a lot of experiments,this paper finds that the input image size has little effect on the accuracy of the model in a certain range,but the smaller the size,the faster the model runs,the less memory usage.Furthermore,we find that the ReLU excitation function is more accurate than the PReLU excitation function,and the operation speed is faster.We also compared the three downsampling methods: Max-Pooling,Lp-Pooling and Convolution,we found that the Max-Pooling has a better advantage in terms of speed and accuracy,and Max-Pooling also has a better filtering effect.This paper presents a real-time semantic segmentation model for unmanned vehicles: FusNet.FusNet uses SegNet's Encoder architecture,and combined with multi-level feature fusion ideas,and on the basis of FusNet we design the two models: FusNet4,FusNet4.Compared with SegNet,the operation speed of FusNet4 and FusNet3 was increased by 9.5 times and 18.9 times,respectively,and in the CamVid database,FusNet4 accuracy increased by 3.05%,FusNet4 has an accuracy of 82.10% on the Cityscapes database.The FusNet presented in this article can provides the real-time processing and the accuracy requirements for smart car.This paper presents a complementary model with FusNet: Deep SemNet.DeepSemNet Combined with semantic edge filtering and Hierarchical ideas,it can be achieved for the task of outdoor scene semantic segmentation.DeepSemNet is a model which built on the basis of PSPNet and EdgeNet,DeepSemNet uses the result of semantic edge detection to filter semantic features.Compared with PSPNet and EdgeNet,DeepSemNet's accuracy increased by 1.76% and 10.5% in the ADE20 K validation set,respectively,and the accuracy of DeepSemNet in the Cityscapes and CamVid are 89.47%,respectively.Compared to the FusNet network which designed in this article,DeepSemNet's accuracy increased by 9.11% and 3.8% in the ADE20 K validation set and CamVid validation set,respectively.