Recognition And Application Of Garbage Cans Overflow Based On Lightweight Convolutional Neural Network

The development of technology has promoted the accelerated construction of smart cities and the initial realization of digital management of cities.With the continuous improvement of living standards,people’s requirements for living environment and quality of life are also improving,and creating a clean and hygienic environment has become an important task in building a smart city.As a common facility in people’s daily life,garbage cans help solve the problem of garbage placement,but if they are not cleaned in time,overflowing garbage cans still bring the bad effect of dirty and messy.Sensor-based detection equipment can make judgments about the capacity of garbage cans and can,to a certain extent,warn of overflowing.However,the installation and maintenance of such detection equipment is a considerable amount of work,and it is impossible to detect the situation of garbage piles.With the development of Deep Learning and the application of Computer Vision,face recognition,vehicle detection and other intelligent services can be realized based on monitoring data.This thesis uses convolutional neural network to build a target detection model to monitor garbage cans based on video images and effectively create a clean and hygienic living environment.In this thesis,the scenarios of overflowing garbage cans are divided into two situations:(1)there is obvious garbage at the garbage can drop-off opening;(2)there is garbage piled up around the garbage can.A relevant dataset was created for both cases.The images were obtained by web download and field shooting,and the data were expanded by color transformation and geometric transformation data augmentation methods.8363 images were obtained,and they were divided into training and validation sets according to the ratio of 8:2.All images are manually labeled to ensure the quality of the data.The algorithm studied in this thesis is mainly used to detect the status of garbage cans in street communities.From the demand that garbage cans should be cleaned in time,a garbage cans overflow detection algorithm based on lightweight convolutional neural network is proposed.In order to improve the performance of the lightweight convolutional model,this thesis proposes a CSA(Coordinate Spatial Attention)module based on the existing attention mechanism.We use Mobilenetv2 network to conduct multi-label classification experiments and verify the effectiveness of the module based on Grad CAM visualization algorithm.In this thesis,we add CSA module to two lightweight target detection models,anchor based and ancho free,respectively,and verify the effectiveness of this module on the detection network through experiments,and obtain the lightweight garbage cans overflow detection algorithm model with the model parametric number of 0.8983 MB and mAP0.5 of 97.0%.In this thesis,the improved lightweight garbage cans overflow detection model is deployed online on the server and the hardware device respectively.The online deployment mainly focuses on The Flask framework and OpenCV are used to build a streaming server to obtain the monitored image data and then load the model for prediction.The terminal deployment is mainly for the scenario without surveillance video,and the hardware device used is PaddlePi-k210.The device comes with a camera and a display,the terminal device does not need to transmit the images to the server,and the detection and recognition of images can be done locally,saving the data transmission generated time and costs.This thesis addresses the problem of lagging garbage cans cleaning in urban governance,creates a relevant dataset and designs and trains an overflow detection algorithm,which can remind relevant staff to clean garbage cans in time and maintain the environmental hygiene.