Research On The Detection Of Itinerant Vendors In Complex Environment

In recent years,the rapid development of object detection technology has greatly promoted the development of autonomous driving,urban safety,intelligent security and other fields.Managing the city well is related to the safety of the people,and how to manage the city is extremely important.Itinerant vendors detection based on object detection results the city management is more efficient and intelligent.Due to the complex background of itinerant vendors,there are situations that affect the detection such as small target scale and severe occlusion,resulting in the existing algorithms cannot be adapted for itinerant vendor detection in complex backgrounds.In response to the above problems,the improvements in this thesis are as follows.The object detection algorithm requires a large dataset for training.A corresponding vendor dataset does not exist,so a itinerant vendor dataset for object detection was constructed by taking pictures of mobile phones and extracting surveillance images.In order to solve the question that vendors are detected toughly in the occlusion environment,a detection algorithm is proposed in this thesis which is based on the algorithm called FPN.The backbone is the residual network with attention module which can reduce the noise weight in the complex environment;The loss function called Focal?Ambig?Loss based on the loss function called Focal Loss is applied to the detect vendors in the occlusion environment.This loss function mainly focuses on objects with indistinct discrimination;FPN uses the mean square error as the loss function of the bounding box regression,and Io U is used as the evaluation standard in the evaluation of positive and negative samples.There is a gap between the two,uses a more reasonable DIo U as the loss function of the bounding box regression.In this thesis,a model of object detection used in detecting vendors called TSF-RCNN was proposed.Which is applied to detect vendors in a long distance.Based on the idea that small objects can be detected more accurately on high-resolution featuremaps.A backbone about three branchs is constructed,every branch is used to detect the specific scale.To keep the sensitivity of the three branches about specific scale,the feature map is selected for scale fusion;Finally,in the step of generating candidate boxes,valid ground truth boxes are selected for each branch.To ensure that invalid boxes do not generate losses,they are all removed.The itinerant vendor dataset set proposed in this thesis is used to verify the effectiveness of the two algorithms.Analysis of the two separate models and the combined model compared with some baselines are carried out.The results show that on the vendors data set,the FPN+SENet model using the improved loss function is 2.5%higher than the basline of FPN+Res Net.And the TSF-RCNN model is 2.7% higher than the basline of Det Net+FPN.Finally,the TSF-RCNN model combines the attention mechanism,the improved loss function Focal?Ambig?Loss and DIo U,which is 3.3%higher than the basline network of Det Net+FPN,and 3.9% higher than the basline network of Det Net+FPN.