Research On The Perspective Method Of Automotive A-Pillar Blind Spot

The A-pillar is the support column closest to the driver among the four beams and columns of the motor vehicle support ceiling.When the car crashes,A-pillar plays A more role in ensuring the life safety of the driver than other support columns.The Apillar is large in size and closest to the driver,thus causing serious obstruction to the driver’s driving line of sight.Therefore,on the basis of not reducing the strength of Apillar,it is of great research value and practical value to solve the problem of blocking the driver’s driving line of sight by A-pillar.The current A-pillar blind zone removal method still has many problems.First,the driver’s eye location algorithm based on deep learning mainly realizes the driver’s eye location according to the image features of the human eye,ignoring the problem that the driver’s eye location cannot be performed when the driver’s face is blocked due to the failure of the driver’s eye image features.Secondly,the existing method does not take into account the problem that the driver’s attention is focused on the front during the driving process,and the factors that may affect the current driving behavior cannot be observed in the A-pillar blind zone.This thesis conducts research on the above issues,and the specific research content is as follows:(1)To solve the problem that the driver’s eyes cannot be located because they are obscured,this thesis proposes a new network structure,YOLOv5 s Contrast,which takes YOLOv5 s as the backbone network and makes improvements on this basis.Firstly,the Rol Pooling layer is introduced into the Prediction department to make the YOLOv5 s network easier to combine with other models.Secondly,based on the idea of comparative learning,the backbone network of YOLOv5 s and Sim CLR network are combined to increase the model’s attention to the spatial information of key points of the driver’s face by randomly masking the input data,so that the model can still realize the ability of positioning the driver’s eyes even when the driver’s eyes are obscured.(2)In order to identify the factors affecting the current driving behavior in the blind zone of pillar A,and then attract the attention of drivers,this thesis proposes a new network structure Separate Deep Sort.In this structure,Deep Sort is the main backbone network,and its feature extraction network is optimized.Firstly,the convolutional neural network in Deep Sort feature extraction network is replaced by deep separable convolution to maintain the same feature extraction capability while reducing the number of convolutional kernel calculations.Secondly,the feature extraction network is improved by introducing residual channel attention network,and different weights are given to different channels to extract important features of the target more effectively.(3)In order to verify the theoretical correctness of YOLOv5 s Contrast and Separate Deep Sort algorithms,this thesis fuses masked face images collected by itself with open face data to obtain a face data set that meets the requirements of training YOLOv5 s Contrast algorithm.Moreover,the ablation experiment and comparison experiment of YOLOv5 s Contrast algorithm were carried out with this data set,and the experimental results were analyzed experimentally,which verified the validity of YOLOv5 s Contrast.In addition,this thesis conducts an experimental analysis of the Separate Deep Sort algorithm on the open data set to verify its improvement in real-time and accuracy.In this thesis,the overall real-time and accuracy of the A-pillar blind zone perspective method is tested,and it is deployed on A solid vehicle to test the actual use effect.After testing,the overall performance meets the actual performance requirements and has certain practical value.