| In recent years,due to the problems existing in the driving process of drivers themselves,the traffic accident rate has increased significantly.To guarantee driving safety and minimize loss when driving,automatic driving has become the key means to solve the problem of traditional driving.The premise of automatic driving is to accurately detect the target on the road.At present,the environment’s complexity affects image identification accuracy,however a target detection method based on depth learning can alleviate the issues in common settings.In this context,under the premise of complex automatic driving scenes,relevant algorithms are studied as follows:(1)The two-stage target detection algorithms Faster Regions with CNN(Faster R-CNN),Single Shot Multi Box Detector(SSD)and You Only Look Once Version(YOLO)are compared and analyzed.An automatic driving data set(ADS)is built for complex automatic driving scenarios.Compare and analyze the ADS data set constructed according to the evaluation indicators.The results show that under difficult conditions,the m AP of yolo4 is 78.55% and the detection speed is 31.84f/s.The accuracy rate is high when the detecting speed is guaranteed.The accuracy rate is high under the assumption that the detection speed will be guaranteed.Consequently,yolo4 serves as the framework for this essay.Further study is required to increase the algorithm’s detection accuracy in complex situations.(2)The factors that affect the detection accuracy in complex scenes are analyzed,and proposed solutions such as integrating attention mechanisms and using improved non maximum suppression algorithms.In order to improve the spatial correlation and channel correlation of small targets,cam attention mechanism is introduced.At the same time,in order to strengthen multi-scale feature fusion,a YOLO prediction head is added to build a 4-scale prediction.CIo U Loss is employed as the regression loss function to increase the precision of the occlusion target position.And the threshold strategy of A-NMS,the attenuation mode of Soft-NMS and the DIo U evaluation mode of DIo U-NMS are integrated into the non maximum suppression algorithm Soft DIo U ANMS.The experimental results show that the improved algorithm achieves 83.24% m AP on ADS dataset,and maintains high detection accuracy on VOC2007 dataset.This demonstrates the enhanced algorithm’s strong generalization capabilities.However,the enhanced algorithm’s detecting speed is just 28.54f/s,which decreases slightly.Therefore,it is necessary to further study this problem.(3)The factors that affect the detection speed in complex scenes are analyzed,and Solutions such as deep separable convolution and lightweight network models are proposed..In order to speed up the detection,Ghost Net is used for feature extraction based on the network structure in Chapter 3.It solves the enormous issues with the original method network model and decreases the amount of parameters that need to be calculated.The ADS dataset is used to test the upgraded algorithm’s detection capacity.According to the results,the revised algorithm’s m AP is 81.83%,and its detection speed is 35.63 f/s.The detection speed is significantly increased on the basis of giving up a tiny amount of accuracy,and the automated driving’s realtime requirements are met.(4)Create a system that can automatically identify driving scenes.Define the system requirements,build the overall scheme,and apply the model that was trained using the enhanced algorithm presented in this research as the system identification model.The ADS data collection is utilized to test the system function.The school’s image is gathered,the various school scenes are captured,and the better algorithm is detected to confirm the algorithm’s detection capacity.This is done in order to validate the improved algorithm’s generalization ability. |
