Research On Occluded Objet Detection Algorithm For Autopilot

In recent years,object detection algorithms based on deep learning have been widely used in autopilot.In this scene,occluded objects are prone to missed detection and misdetection,which seriously affects the driving safety of autonomous vehicles.At present,most of the occlusion problems are dealt with anchor-based two-stage object detection algorithms.The disadvantages are that the detection speed of algorithm model is slow and the hyperparameter selection is complicated.In addition,for the occluded detection object,the object features extracted by the deep neural network are relatively incomplete and lack the refinement of the feature level.To solve these two problems,this thesis proposes to combine feature fusion and attention mechanism on the basis of the anchor-free object detection algorithm(CenterNet algorithm)to realize the detection of occluded objecs in autopilot.The main work of this thesis are as follows:1.Firstly,explain the development and advantages of deep convolutional neural networks,and analyze the anchor-based representative detection algorithms(Faster R-CNN,YOLO and SSD)and the anchor-free representative detection algorithms(CornerNet,FCOS and CenterNet).Then combine with the autopilot scene,summarize and analyze the three methods commonly used to deal with occlusion problems(component optimization,loss function optimization and feature fusion methods),and provide a theoretical basis for the subsequent improvement methods.2.The CB-CenterNet(Cascade Backbone-CenterNet)algorithm is designed based on the CenterNet algorithm combined with the idea of differential cascade feature fusion,which the enhanced composite connection module used to perform differential cascading feature fusion of two DLA34 backbone networks to form CB-DLA34 feature extraction network.In the detection output sub-network,the gaussian heatmap is generated using the improved gaussian kernel function with adaptive control of gaussian radius,and the regression samples are sampled within its gaussian region.The experimental results show that compared with the current mainstream detection algorithms,the CB-CenterNet algorithm improves the m AP(mean Average Precision)by 0.8%～15.8% on the selected BDD100 K dataset,which increased by 2.9% compared with the CenterNet algorithm,and training time reduced by 21.2hours.3.The Dense CB-CenterNet algorithm is designed based on the CB-CenterNet algorithm combined with the densely connected attention mechanism,and of which the Dense-CBAM(Convolution Block Attention Module)densely connected attention module with cross-dimensional interaction is embedded among different stages of the feature extraction network to form the Desnse CB-DLA34 feature extraction network.This thesis draw the idea of Repulsion Loss to deal with the occlusion problem,and design R-CIo U Loss with attractive loss and repulsive loss based on the bounding box regression loss CIo U Loss.The experimental results show that compared with the CB-CenterNet algorithm,the m AP(mean average precision)of Dense CB-CenterNet algorithm has increased by 1.5% on the selected BDD100 K dataset,and improved by 4.4% compared with the CenterNet algorithm.