Research On Semantic Segmentation Algorithm Of Parking Lot Vehicles Based On Fully Convolutional Neural Network

With the arrival of China's automobile era,‘Parking hard' is a big problem that needs to be solved urgently,It is an efficient and low-cost solution to accurately identify and locate the vehicles and empty spaces in the parking lot by using the existing cameras in the parking lot.In recent years,high-performance computing developers represented by NVIDIA have led the leapfrog development of computing power;convolutional neural network dominates the whole image field and it has achieved a brilliant result.The semantic segmentation of the image locates the vehicles in the parking lot with pixel-level accuracy,moreover,the emergence of the Full Convolutional Neural Networks(FCNs)provide a new driving force for instance semantic segmentation;however,the lack of large-scale semantic image data sets greatly weakens the generalization ability of neural networks;the downsampling of deep convolutional neural network and the sparsity of feature map will lead to the loss of small-size targets;parameters of high-quality instance semantic segmentation network are numerous,and it is easy to overfit during training.In the face of the above problems,this paper studies the semantic segmentation of vehicles in the parking lot.,its main work is as follows:To improve the generalization ability of the instance semantic segmentation network,we introduce the residual network(ResNet)and large-scale increase of semantic image data.In collecting data sets,we mainly start from two aspects,first,start with real image data,unify the current mainstream open source dataset(such as: MS COCO,PASCAL VOC,SUN,Cityscapes,etc.),extract vehicle data,and create a new data set according to MS COCO standards;second,start with virtual data,based on Blender 3D rendering engine,making realistic virtual parking lot and vehicle image,combining with edge detection algorithm to achieve high precision labeling of images.Aiming at the sparsity of the feature map,we propose a feature concatenation scheme,the shared features well incorporate deep but highly semantic,intermediate but really complementary,and shallow but naturally high-resolution features of the image,In the process of feature concatenation,we have abandoned the maxpooling and Bilinear interpolation schemes for feature map sampling,but introduced the dilated convolution with learning ability and deconvolution(Deconv)to sample the feature map.Based on the MS COCO dataset,feature concatenation achieves an increase of 0.5% in the mAP of the instance semantic segmentation network.In view of the problems caused by downsampling of deep convolutional neural networks,we propose a Dual RPN network based on RPN and a new data argumentation scheme.Dual RPN network is a combination of shallow RPN and deep RPN through soft-NMS,in the process of implementing this scheme,we propose a semantic enhancement network to solve the problem of poor classification performance of shallow RPN networks.Based on the PASCAL VOC dataset,when the IoU is 0.5,0.6,and 0.7,respectively,the Top-1000 predictive score bboxes are counted,The average recall of the Dual RPN is 3.2%,8% and 10.5% higher than that of the RPN,respectively.Based on the MS COCO dataset,Dual RPN and data argumentation increased the mAP of the instance semantic segmentation network by 0.4% and 2.5%,respectively.For algorithm implementation and training issues,because parameters the semantic segmentation network proposed in this paper are numerous,constructing it directly into an end-to-end network is not conducive to parameter tuning,here we use step-by-step training to train the sub-network to the optimal,and then integrate the sub-network model parameters to fine-tuning the entire network.