Research For Real-time Multi-class Vehicle Detection Technology Based On Computer Vision

In recent years,with the rapid development of urban transportation,intelligent transportation systems have also received more and more attention,and the completion of real-time vehicle detection and category recognition is the most important technology in intelligent transportation systems,and it is also one of the key issues in the current intelligent transportation field.At the same time,the research of vehicle detection is also of great significance for improving traffic control and alleviating traffic pressure.The methods in this field require high robustness,high-scale insensitivity,real-time and high precision.Traditional vehicle detection algorithms have encountered difficulties in solving high-precision and real-time performance.Meanwhile,the detection algorithms based on deep learning that were born in recent years have some scale-sensitivity problems,especially for small object detection.This article researches and analyzes the above problems and proposes two novel vehicle detection schemes.(1)A novel multi-category vehicle detection algorithm based on MOG2 and HSqueeze Net is designed.First,the traditional background modeling method MOG2 is used to complete the regional proposal task and find the location of the vehicle.Then,the paper designs a model named H-Squeeze Net to identify the vehicle category,based on the deep learning network model Squeeze Net.While completing the vehicle detection task,the speed is achieved 39.1FPS,which meets the real-time requirements.In order to verify the effectiveness of the proposed algorithm,this paper evaluated the algorithm on the CDnet2014 dataset and verified its practical application on the collected data of Suzhou Huqiu traffic intersections.Aiming at the scale sensitivity problem in deep learning algorithms,this algorithm introduces the MOG2 algorithm to the regional proposal field,and uses it to generate robust scale-insensitivity regional proposals.At the same time,in order to obtain high precision and high performance,the algorithm uses H-Squeeze Net to quickly identify vehicle category in the backend.The proposed algorithm can avoid each other’s shortcomings as cleverly as possible while retaining the advantages of both.(2)We propose a novel YoLoV3 detection scheme YoLoV3＿Squeeze Net based on Squeeze Net.This scheme uses the single-stage detection algorithm YoLoV3 based on deep learning as the framework,improves its feature extraction model,removes its Dark Net53 model,and tries to utilize the lightweight network Squeeze Net as its feature extraction model.During the experiment process,it is difficult for Squeeze Net to extract suitable feature maps for YoLoV3 to train and utilize.This paper introduces deep separable convolution to transform the extracted feature maps,and successfully migrates Squeeze Net to YoLoV3.Meanwhile,the application of the lightweight network will greatly reduce the amount of model parameters and the size of the weight model at small cost of accuracy.The weight model size after migration is only 36.01%of the original model,and it can speed up 25.01% of model inference time and 15.3%of training time.Moreover,the paper also introduces the UA-DETRAC traffic dataset to evaluate YoLoV3＿Squeeze Net.The performance shows that the algorithm can achieve better vehicle detection results in different weather and road conditions.In summary,the paper utilizes road traffic dataset as the research foundation,vehicle detection as the research goal,real-time performance and high performance as the research goal.By building a real-time multi-category vehicle detection scheme,and training and testing on the traffic dataset,the vehicle detection task is finally realized.