Research On Non-motor Vehicles Illegal Parking Detection Method Based On Deep Learning

With the rapid development of China’s economy and the improvement of people’s living standards,the transportation means used by publics are becoming more and more diversified.More and more motor vehicles and non-motor vehicles play an important role in travel,and the ensuing congestion is also more and more frequent in people’s daily life.In urban life,the indiscriminate parking of electric motorcycles and shared bikes will bring the risk of road traffic congestion.At the present stage,the method to deal with illegal parking of non-motor vehicles is mainly through manual inspection,through the investment of a large number of traffic police and city management staffs,inspection street by street and detection of the phenomenon of illegal parking,so as to eliminate the risk of causing congestion.At present,there are many defects in manual detection,which is laborious and time-consuming,inability to continuously detect,and misjudge or miss detection due to tiredness.Compared with the conventional manual detection methods,the non-motor vehicle parking violation detection method based on deep learning has the following advantages:first of all,in the whole detection system,the corresponding street can be detected in real time and uninterrupted by the deployed image acquisition equipment;secondly,In the illegal parking detection,the detailed information of illegal parking can be reported in real time.Once a serious illegal parking phenomenon occurs,a warning will be issued to the system immediately,which can be used by road enforcement officers for reference.The research on illegal parking detection of non-motor vehicles can not only detect and rectify illegal parking on the road in time,but also alleviate road congestion and further the road conditions.The main research content of this paper is non-motor vehicle parking violation detection based on deep learning,and the research combines cutting-edge technologies such as weakly supervised target positioning and self-supervised learning.The main work and innovations of the paper are as follows:(1)A method for detecting illegal parking of non-motor vehicles in street view images is proposed.This method uses the fully supervised object detection algorithm and the semantic segmentation algorithm to obtain the non-motor vehicle target detection information and pixel-level semantic information in the image to judge whether there is illegal parking of non-motor vehicles in the street view image.(2)An improved adaptive attention augmentor weakly-supervised localization method is proposed.Through weakly supervised learning,non-motor vehicle object localization can be performed when only image class labels are provided.This method obtains more accurate target regions by using an adaptive correction module and applying both spatial attention mechanism and channel attention mechanism.Compared with the original adaptive weakly supervised localization method A~3,the method proposed in this chapter can obtain more complete target regions and higher classification accuracy,and has better visual effects.(3)A self-supervised representation learning algorithm based on LEGO sampling is proposed.Through self-supervised learning,intra-features in images can be extracted without labels and applied in downstream non-motor vehicle object detection tasks.The method first introduces a LEGO sampling strategy to increase the number of samples by sampling smaller patches in the original image,while using the image stitched with smaller patches to maintain less computation,and secondly introduces a local detail contrast branch to balance local details The relationship between the features and the global semantic features,and finally use a variety of loss functions to jointly optimize the model.Compared with other existing self-supervised learning methods,the presented approach in this thesis can obtain more comprehensive global information and local details,have better visual effects,obtain better classification accuracy and average detection accuracy in downstream linear classification,target detection and other tasks.