Research On Key Technologies Of Automotive Active Safety System Based On Computer Vision And Deep Learning

In the process of intelligent development of automotive,the craze of automatic driving has swept the world.However,Sebastian Thrun pointed out that automatic driving is still at an early stage and will not be widely spread in the short term.It will be 8 to 10 years before driverless cars can be used by most people in their daily lives.Semi-automatic driving based on intelligent and active safety technology is more practical.Automotive safety is the focus of automotive intellectualization.Nowadays,the automation and artificial intelligence technology combined with big data from Internet of Vehicles are promoting the development of automotive industry from the Passive Safety to Active Safety in depth and breadth.In recent years,research on key technologies of Automotive Active Safety System(AASS)based on computer vision and deep learning has become a hotspot and frontier in this field,which has an important social significance and application value to improve driving safety and reduce traffic accident rate.Computer vision and deep learning theory are taken as technical means,and the AASS is taken as research object in this study.Some key technologies of AASS,such as environmental perception technology(object detection and tracking),driving behavior analysis technology(lane change behavior detection and prediction),and intelligent control and decision technology(intelligent collision avoidance control strategy)have been carried out,and the experimental results are presented.The main innovations of the research are as follow:(1)In order to enhance the ability of environmental perception,the research on moving object detection and tracking under driving environment have been studied.In order to address the problem of traditional object tracking algorithms,such as unable to adapt to complex tracking changes and the low accuracy when tracking target is occluded.A moving object tracking algorithm based on You Only Look Once(YOLO)and Recurrent Regression Network(RRN)is presented,which is a simple,fast and accurate recurrent regression network model.The model meets the real-time requirement of environmental perception well,and has strong robustness to object tracking with occlusion,which can provide effective guarantee for tracking performance of AASS in complex environment.(2)On the basis of environmental perception,a sensor fusion approach is proposed that utilizes two differing modality data,that is,road view video data and On-Board Diagnostics data to extract features,in order to enhance the ability of lane change behavior detection under driving environment.Two fusion methods(feature-level fusion method and decision-level fusion method)are verified using a novel model,Collaborative Representation Optimized Projection Classifier(CROPC).The experimental results showed that,in comparison to other state-of-the-art classifiers,CROPC model performs significantly better,which can provide higher accuracy for lane chang detection under driving environment.(3)Prediction of lane change behavior is the key to realize intelligent collision avoidance control.In order to predict driver's lane change intention before it actually occured,a novel Multivariate Time Series Group-wise Convolutional Neural Network(MTS-GCNN)prediction model is proposed.In the MTS-GCNN model,a new structure learning algorithm is presented in training stage.The algorithm exploits the covariance structure over multiple time series to partition input volume into groups,then learns the MTS-GCNN structure explicitly by clustering input sequences with spectral clustering.The MTS-GCNN can select and extract the suitable internal structure to generate temporal and spatial features automatically by using convolution and down-sample operations.Therefore,a fast and reliable automatically lane change prediction process is implemented.(4)A novel Deep Discriminant Model(DDM)is proposed for predicting imminent collisions caused by dangerous lane change and collision risk assessment,which lay the foundation for intelligent collision avoidance control task of AASS.In particular,a special network,ConvLSTMs is presented,which is a combination of convolutional and recurrent layers,to help the DDM processing visual data,vehicle status data and physiological signal data in spatial-temporal domain simultaneously.The experimental results showed that,it is feasible and beneficial to process multiple modalities of data in a deep learning framework.A new idea is provided for the research on automotive active safety system in this study.The proposed research methods and experimental results have proved that computer vision technology and deep learning model have an important theoretical significance and potential application value in the field of automotive active safety.