Application Of Pedestrians Detection Algorithm Based On Convolutional Neural Network In Passengers Counting

In recent years,the continuous progress of Internet technology has greatly promoted the development of e-commerce.More and more people are keen on online consumption,which has brought huge impact to offline brick-and-mortar stores.As a type of operation of physical stores,convenience stores play a very important role in people's life and are also influenced by e-commerce.The profit of convenience stores is directly related to the flow of customer.Through the statistics of customer flow,the scientific analysis of customer flow data and the formulation of a reasonable marketing plan are of great significance to cope with the challenges brought by e-commerce and achieve considerable development.Pedestrians detection technology is very pivotal to locate accurately pedestrians in images or videos,which provides important technical support for video surveillance,assisted driving,autonomous driving,intelligent robots and many other fields as well.Because of the wide application of pedestrians detection,it has become an important research topic.In this thesis,the computer vision technology is used for counting the passengers of convenience stores,and pedestrian detection is the key of the system.The work of this thesis is as follows:(1)Collected pedestrians datasets.MS COCO is currently the largest and most authoritative object detection dataset with a total of 80 object categories,including 328 k images.For the purpose of study,I extracted all data that containing people out of the MS COCO,which consist of 66 808 pictures,and dividing randomly them into training set,verification set and test set in a ratio of 2:1:1.(2)Considering the accuracy and processing speed of pedestrian detection model,the YOLOv3 Tiny algorithm was improved which based on convolutional neural network.In this thesis,the YOLOv3 Tiny algorithm is reproduced on the collected pedestrians dataset.To solve the problem of low accuracy on this dataset,the original algorithm is improved in four methods.The first,VGG16 model has shown good feature extraction ability in a variety of computer vision tasks,so it was applied as frontend in YOLOv3 Tiny network architecture.The second,the original algorithm was redesigned with the idea of residual network.The third,CNN need to extract more meaningful advanced features by downsampling,but the operation will lead to loss of information inevitably.For solving the flaw,I put forward the third improvement strategy.That is using 2 strides and 2 dilated rate convolution to downsample,expanding receptive field at the same time.The last version,this model aim to design a lightweight model to quicken runing as much as possible.Experimental results show that the third improved method has the best performance on the collected pedestrian dataset,and the accuracy is improved by 12.3% compared with the original algorithm,and the processing speed reaches 61 FPS on RTX2080.(3)Developed a passengers counting system for convenience store.The system carries out real-time detection on the pedestrians in the surveillance video,and judges whether the pedestrians pass through the entrance of the convenience store through the Ray Casting,so as to count the customers of the convenience store.