Font Size: a A A

Fixed-Angle Road Rubbish Detection Based On Convolutional Neural Networks

Posted on:2020-09-06Degree:MasterType:Thesis
Country:ChinaCandidate:X P YuFull Text:PDF
GTID:2428330590958267Subject:Control Science and Engineering
Abstract/Summary:PDF Full Text Request
With the development of modern society,people's living standards are increasing,and the amount of rubbish is increasing,too.Urban road often sees a variety of rubbish,which seriously affects the urban environment.Moreover,the main way to protect urban environment now is manual regular inspection and cleaning,which is inefficient and impossible to find the rubbish and clean the road in time.At present,urban roads are equipped with surveillance cameras.Therefore,road surveillance cameras or fixed-angle cameras can be used to obtain road images,real-time detection of road rubbish and notification of relevant personnel cleaning,and improve the level of urban environmental protection intelligence.Since road rubbish distribution has no time and space rules,various types,rubbish definition is difficult,and road rubbish has fewer object pixels in the image,which brings great challenges to the detection of road rubbish.In this paper,a spatially constrained road rubbish detection algorithm based on convolutional neural network is proposed.Real-time detection of rubbish on the road image acquired in real time and combined with the road area detected by the clean road background image from the same perspective angle as the space constraint to detect the road rubbish.It solves the difficulty of road rubbish definition and small road rubbish object detection.The method mainly includes two stages: road area detection based on improved full convolutional network(FCN)and small rubbish object detection based on improved SSD.The main contents of this paper are as follows:Aiming to the unstructured roads and complex scenes in road detection,this paper proposes a road detection algorithm based on improved full convolutional network,which improves the ResNet-50 feature network,and adds dilated convolution and setting in some convolutional layers.The smaller sampling step size is used to obtain the feature map of the larger receptive field and resolution.The improved ResNe50 is used as the basic feature extract network of the FCN,and a multi-level upsampling module is designed to classify the road image by pixel level,detecting road areas.Experiments show that the road detection algorithm has good performance,and achieves 96.786% pixel accuracy and 93.527% MeanIOU on the road detection data set.Targeting at the difficulty of small object detection of road rubbish,this paper proposes a road rubbish detection algorithm based on SSD object detection framework,which integrates the feature layers of different adjacent three layers in SSD to improve the semantic of shallow network layer;According to the distribution of the height and width of road rubbish object,the aspect ratio and the number of the default boxes of the different layers of the SSD is designed.Experiments show that the context feature fusion module and Default Box design of SSD can be good used for road rubbish detection.In order to further improve the performance of road rubbish detection,based on the idea of channel selection of the SENet and ResNet residual connection,a feature enhancement model for SSD is proposed,and two different general feature enhancement modules are designed.Experiments show that the feature enhancement module can improve road rubbish detection performance.At the same time,the combination of feature enhancement with upper and lower feature fusion modules can further improve the performance of SSD road rubbish small object detection,achieves the 81.311% mAP in the dataset of this paper and greatly solves the problem of road rubbish detection.
Keywords/Search Tags:Road rubbish detection, Object detection, Convolutional neural network, SSD, Fully Connected Network
PDF Full Text Request
Related items