Research On Deep Learning-based Object Detection Algorithm For Autonomous Driving

With the rapid development of artificial intelligence technology,autonomous driving has become a strategic direction for the development of the global automotive industry.Object detection,as an important component of the autonomous driving perception system,has a direct impact on the subsequent decision-making behavior and the safety performance of autonomous driving.Due to the poor feature generalization ability and complex algorithm design,traditional object detection algorithms are difficult to achieve the accuracy and speed requirements of autonomous driving.Deep learning-based object detection algorithms can effectively improve the above deficiencies and meet the autonomous driving perception requirements.For autonomous driving perception needs,this paper improves the deep learning-based two-stage object detection algorithm Faster-RCNN and the one-stage object detection algorithm YOLOv3,and validates the algorithm performance on a large autonomous driving dataset BDD100 K.The main contents of the study are as follows:For the inaccurate localization and missed detection obscured target of in complex traffic scenes,the localization loss function and bounding box postprocessing algorithm of the two-stage object detection algorithm Faster-RCNN are studied,and a complex scene object detection algorithm based on improved FasterRCNN is proposed.To improve the localization accuracy of the occluded target,the overlap degree and position relationship between the predicted bounding box and the ground truth box are introduced in the localization loss function,and the CIo U Loss localization function is used to improve the Faster-RCNN algorithm.To alleviate the missed detection problem of the occluded target,the deficiencies of the NMS are analyzed,and the Soft-CIo U-NMS bounding box post-processing algorithm is proposed.The experimental results show that the improved Faster-RCNN improves the Precision,Recall and mean average precision(m AP)by 1.1%,2.5% and 3.0% on the autonomous driving dataset BDD100 K,which effectively alleviates the problem of inaccurate localization and missed detection of obscured targets in complex traffic scenes.For the problem of multi-scale object detection such as vehicles,pedestrians,and traffic signs in the autonomous driving environment,the anchor generation method and the fusion method of multi-scale features of YOLOv3 are improved,and the multiscale object detection algorithm based on the improved attentional feature pyramid is proposed.To improve the quality of anchor,the bounding boxes in the dataset are clustered,and the anchor generation method based on K-Means and genetic algorithm is proposed.To improve the multi-scale object detection accuracy,the shortcomings of the feature pyramid network are analyzed,a reasonable multi-scale feature fusion method is designed,and the attention feature pyramid network is proposed.The experimental results show that the attention feature pyramid network improves the m AP of large,medium and small targets of different scales by 0.9%,1.1% and 1.0%with a small number of parameters increase,which effectively improves the detection of multi-scale targets.For the real-time deployment of autonomous driving object detection,the lightweight design method of YOLOv3 is studied and a lightweight object detection algorithm based on dual-blueprint convolutional network is proposed.Analyzing the redundancy of feature extraction network Dark Net53 to obtain cosine similarity of convolutional kernels and principal component analysis results of convolutional channels.To alleviate the convolutional kernel redundancy and channel redundancy of Dark Net53,a dual-blueprint convolutional network is proposed to make the model more lightweight.The experimental results show that the model parameters of the lightweight object detection algorithm based on the dual-blueprint convolutional network are reduced by 27.18 M,the detection speed is improved by 3fps,and the detection accuracy m AP is improved by 2.3% compared with YOLOv3.Compared with the improved Faster R-CNN,the algorithm meets the real-time requirements of autonomous driving while holding a higher detection accuracy,and is more suitable for the scenario of autonomous driving object detection.