Research On Vehicle-Mounted Video Detection And Tracking Algorithm Based On Raspberry PI Platform

With the increase of the number of cars,traffic safety has become an important issue.As an effective means to prevent traffic accidents,the importance of relevant assisted driving technology is self-evident.Lane lines and obstacles,as the most basic elements of assisted driving technology,are the most important.At present,the research on assisted driving technology can be divided into two research directions.One is based on the Internet of vehicles,but there are hidden dangers of communication delay or interruption;In addition,the research direction of this thesis is related to the research of assisted driving technology based on the embedded platform.The embedded platform is easy to install in the vehicle,and there is no communication interruption problem,which is more guaranteed.Based on the embedded raspberry PI platform which is small in size,low in cost and easy to transplant,this thesis studies the algorithm of lane line detection and dynamic obstacle detection and tracking through vehicle-mounted video.The main research contents of this thesis are as follows:Aiming at the poor real-time performance of traditional lane line detection algorithm on embedded platform,a fast lane line detection algorithm is designed in this thesis.In the image preprocessing stage,an adaptive binarization extraction algorithm of lane lines was designed.By comparing the pixels to be measured with the vertices of the rhomboid space,lane line information was extracted quickly and completely.At the same time,combined with the maximum variance between classes(OTSU),the interference information can be effectively filtered by image fusion.In the lane line fitting stage,the slope constraint and limited distance of probabilistic Hough transform are improved,and the lane line edge points are calculated quickly and accurately after secondary filtering of interference information.Finally,the least square method is used to fit the lane lines.Experimental results show that the average accuracy of the algorithm can reach 90.24% on raspberry PI platform,and the running speed is 25 fps,which can run in real time.Aiming at the problem that the false detection rate was too high due to the easy ghosting of Vi Be algorithm when dynamic obstacle detection was carried out on the raspberry PI platform,Vi Be algorithm was improved.Firstly,a time background model is added for each pixel based on the time characteristic of the pixel and a certain initialization time is given.Secondly,aiming at the problem of ghosting caused by moving objects in the image when Vi Be algorithm used the first frame to build the spatial background model,a background replacement method was designed,which could effectively replace the moving objects with the real background.Then,the segmentation threshold of the traditional Vi Be algorithm is adjusted adaptively,and a foreground detection strategy is designed combining with the time background model,which can reduce the noise interference and make the ghost quickly dissipate.Finally,an update strategy is designed to update the two background models quickly and stably.The improved algorithm has better suppression of the ghost,and the accuracy rate is increased by nearly 24%.It can complete the dynamic obstacle detection task well on the raspberry PI platform.In order to solve the problem that the scale of tracking frame is fixed and the ability to deal with occlusion and loss condition of tracking target is poor when using KCF to track obstacles,the KCF algorithm is improved.First,the color features of CN were introduced and the feature fusion was carried out to enhance the target’s feature expression ability.Secondly,the scale pool is introduced and divided.After dividing the scale pool,an independent one-dimensional scale filter is trained to predict the scale,which can better adapt to the situation of scale mutation on the premise of guaranteeing the speed.Finally,the model updating mechanism is improved and the re-detection mechanism is proposed.The tracking confidence index is combined with the average peak correlation energy and maximum response value to distinguish whether the target is in abnormal state,so as to judge whether the model is updated or not.At the same time,sidelobe detection is used to further subdivide the status to determine whether to enable the redetection step.The combination of the two mechanisms can effectively improve the model drift phenomenon caused by target occlusion and loss.The test results on raspberry PI platform show that the improved algorithm has higher accuracy and success rate,and can guarantee higher running speed.