Research On Object Detection And Tracking Technology In Automotive Vision

With the rapid development of the national economy,the number of car ownership has increased year by year,which has brought about a series of traffic safety problems while facilitating people's daily life.In order to improve driving safety and reduce traffic accidents,driver assistance systems have emerged.A safe and reliable driver assistance system usually needs to connect to the internet.However,the traditional Internet has the problems of low communication efficiency and long transmission delay.The dual-structural network advocates integrating the current internet architecture with a broadcast-storage secondary structural network characterized by “radiation-copy” model,which can effectively reduce the traffic load of the internet and latency.At the same time,with the help of Uniform Content Label(UCL)which contains rich semantic features,the driver assistance system can realize the standardized processing of object information and realize secure network communication.Driver assistance systems require a lot of technical support,object detection and tracking is one of the key technologies of the system.However,the reality scene is complex and changeable,it is difficult to detect and track the object accurately and efficiently.In the object detection task,the SSD detection algorithm has been widely studied and applied due to its real-time detection speed and high detection accuracy.However,this algorithm has a high rate of missed detection of small objects.In the object tracking task,a single object tracking algorithm is difficult to solve the above problems in the tracking task,and the tracking effect is usually poor.Aiming at solving the problems above,this dissertation proposes a BSSD object detection algorithm based on feature fusion which named Bidirection Single Shot multibox Detector,and puts forward an object tracking algorithm BSSD-MSE based on BSSD detection and motion state estimation which named Bidirection Single Shot multibox Detector and Motion State Estimation.On this basis,this dissertation designs an object detection and tracking system for automotive vision to verify the algorithm proposed in this dissertation through experiments.The main work of this dissertation is as follows:(1)Aiming at the problem that traditional SSD algorithm is difficult to detect small objects,this dissertation puts forward a BSSD object detection algorithm based on feature fusion.Firstly,based on the feature fusion in FPN and the passthrough theory in YOLOV2,this dissertation constructs an information-rich feature layer by merging the feature layers of different scales,and uses this feature layer to detect small object.Then this dessertaion improves the data augmentation strategy of the traditional SSD algorithm,adding the small sampling ratio and propose a new data augmentation strategy to enrich the small object training sample set.Then this dissertation integrates Focal Loss into the loss function to reduce the impact of data imbalance on the model.Finally,performing UCL indexing on the detected object information to facilitate target tracking module to extract the detected target fastly,and reporting the indexing result,distributing it to other vehicle terminals by broadcasting.(2)Aiming at the problem that the single object tracking algorithm is difficult to deal with the complex and variable tracking environment,this dissertation puts forward an object tracking algorithm BSSD-MSE based on BSSD detection and motion state estimation.Firstly,this dissertation initializes the tracking object by using BSSD object detection algorithm and ResNet-18 classification algorithm.Then this dissertation divides the state of the object,using Kalman filter algorithm and kernel correlation filter algorithm comprehensively to estimate the motion state of objects in different states.Finally,according to the correlation matrix,this dissertation associates the motion state estimation results and the detection results,and convert the state of the object,in order to achieve tracking of the object.Finally,using UCL to encapsulate tracking object information and report encapsulation results,in order to assist other vehicle terminals to make correct driving decisions.(3)Combined with the characteristics of dual-structural network,this dissertation realizes the object detection and tracking prototype system for automotive vision.This dissertation tests and analyzes the BSSD object detection algorithm and BSSD-MSE object tracking algorithm by experiments.The experimental results show that the BSSD algorithm has higher mean average precision and racall rate than the traditional SSD algorithm,and can detect more small objects.The BSSD-MSE algorithm can cope with the complex tracking environment goodly,and the tracking speed of the algorithm is fast and the robustness is good.BSSD-MSE algorithm can achieve the long-term and accurate tracking of the objects.