Research On Autopilot Occlusion Object Detection Algorithm Based On Deep Learning

With the continuous development of deep learning technology,object detection algorithms based on deep learning are also changing rapidly and their detection performance is getting better and better.It can be used in environment perception links in autonomous driving to detect traffic participants.However,traffic participants blocking each other will cause false or even missed detection,which seriously affects the safety of autonomous driving.This thesis studies the occlusion target detection algorithm of autonomous driving occlusion,and detects the four object in this scene.Based on the convolutional neural network and Transformer in deep learning technology,the modified Mask RCNN and SRA＿Swin Mask RCNN object detection algorithms are designed to realize the object detection of autonomous driving occlusion,respectively.The main work is as follows:1.Build an image dataset of four object in autonomous driving scenarios.This dataset is chosen in BDD100 K,include 16591 images,including 4 categories: car,bus,pedestrian and truck,of which 12923 images serve as train set and 3668 images serve as test set.In addition,the mild occlusion,moderate occlusion and severe occlusion test sets is made for different degrees of occlusion,with 65 images in each degree of occlusion test set.2.An autonomous driving occlusion object detection algorithm based on an improved Mask RCNN network is designed and implemented.Non-overlapping convolutions of 4×4 and the convolution structures of 1×1 is designed in the residual structures in the preprocessed part and the backbone,respectively.The modified Mask RCNN network is also designed using repulsion loss as the bounding box loss function of Mask RCNN.The experimental results show that the m AP(mean Average Precision)of improved Mask RCNN model is improved by 2.82%compare to the original Mask RCNN model,especially the m AP of detect person by 4.53%.The detection accuracy of no occlusion and different degrees of occlusion has been improved.3.The autopilot occlusion object detection algorithm based on SRA＿Swin Mask RCNN is designed and implemented.It consists of backbone SRA＿Swin Transformer,Region Proposal Network,ROI Align module and loss function.Spatial-Reduction Attetion and the repulsion loss bounding box regression loss function are used.The experimental results show that compared with the Mask RCNN original model,the m AP of the four targets is improved by11.3%;Compared with the improved Mask RCNN model,the SRA＿Swin Mask RCNN model improved the m AP by 8.48% for the four objects.