Research On Vehicle Detection Algorithm Based On Image Assessment And Optimization

The vehicle detection system is an indispensable part of the intelligent transportation system and it is great significance to guide the vehicle diversion scientifically and maintaining the safety of public transportation.At present,the vehicle detection technology based on computer vision has been widely used.The video surveillance system can monitor all-round vehicle on the road.The massive collected videos of vehicle on the road can as the database for the research of vehicle detection technology.However,in actual situations,the quality of collected vehicle images is uneven cause by complicated weather,different parameters of various monitoring equipment,vehicle speed difference,etc.The low-quality images can degrade the learning ability of the deep learning network and reduce vehicle detection accuracy.In order to obtain ideal identification results of mult-object under various circumstances,researchers generally select deeper network with higher detection accuracy to set up general object detection model at the expense of running time.It is difficult to meet the real-time requirements in vehicle detection applications.In order to solve the above problems in vehicle detection technology,this paper proposes a vehicle detection algorithm based on image assessment and optimization for the vehicle detection in complicated environments.The proposed method based on the model compression technology and channel attention mechanism that can improve vehicle detection speed on the basis of ensuring the detection accuracy.This paper is summarized as follows:Firstly,for motion blur vehicle images,this paper proposes a sort-based motion blur image quality assessment algorithm to filter out motion blur images.The purpose of the algorithm is provide guidance for the next step of vehicle image optimization and detection.On the one hand,we use the collected high-quality images to create a database of motion blur images for model training.On the other hand,this paper designs an image quality ranking model and an image quality assessment model based on the twin network structure,and uses a scoring mechanism based on the sliding window method to accurately evaluate the quality of the image.Secondly,for a single category detection tasks of the vehicles,a vehicle detection algorithm CA-Net based on feature recalibration is designed.On the one hand,according to the visualization results of feature map,we introduce channel attention mechanism to improve the convolutional structure and lightweight YOLOv3 for improving detection speed.On the other hand,we use multi-scale prediction mechanism to detect and identiy vehicles with various sizes.The experiment results prove that the vehicle detection accuracy of the CA-Net model achieves 93% and the detection speed reaches 57 FPS.Finally,in order to alleviate the vehicle missed detection cause by motion blur images,this paper combines image quality assessment and optimization with vehicle detection algorithms to design a vehicle detection algorithm.The processing can be devided into training and testing.In the training,we evaluate and optimize the vehicle detection database by the image quality evaluation algorithm and image optimization algorithm to prepare the high-quality vehicle database for training the vehicle detection model.In the testing,we assess motion blur degree of the testing vehicle images,and then optimize the motion blur image according to the assessment result.We send the image without motion blur to the model for detection.The Experiment results show that the proposed algorithm can improve the detection accuracy of motion blur vehicle images from 87% to 92%.