Obstacle Detection Under Deep Learning

Due to the vigorous development of artificial intelligence technology,people have put forward a series of indicators such as high precision,high performance,and strong adaptability to the environment for unmanned driving technology.However,road obstacle detection is still an insurmountable barrier in the field of visual intelligent driving.It can be said that improving the accuracy of road obstacle detection can greatly reduce the false detection rate of intelligent driving,and indirectly provide drivers with a safe and comfortable driving or riding environment..This technology relies extensively on computer vision-related technologies to automatically detect various obstacles on the road and provide accurate and reliable information to provide important support and guarantee for the driving of autonomous vehicles.In recent years,a large number of typical research results have emerged in the direction of obstacle detection.Although there are many solutions for road obstacle detection,the following problems still need our attention: the road information is changing,the flow of people and traffic is large,and the background information is rich and huge,which directly leads to longer extraction time of image features,and the influence of weather and light.It is difficult to divide the obstacle detection frame,and the size and shape of obstacles make the obstacle detection system less accurate.This paper proposes a new network algorithm structure for these problems: this paper integrates and updates the VOC2007 data set and the Ex Dark data set to improve the ability of the model to adapt to different environments such as low light,strong light,and night.Inspired by the residual network,this paper Improve the network system of Faster-RCNN,and create a network structure of multi-layer feature information fusion to stack the residual network BTNK1 and BTNK2 to solve the phenomenon of long extraction time of some obstacle detection features in complex environments.At the same time,integrate Momentum and Ada Grad to create Momentum-Ada Grad as a regression detection frame optimizer,and use the softmax function as the activation function of the fully connected layer to improve the detection ability of the network model,and improve the training speed by freezing and thawing training.Finally,the detection effect of the network in this paper is verified by comparing with the traditional network structure.This paper proposes an improved method of the Efficient Det network model,and improves the detection accuracy on the basis of Efficient Det-B0.Normalize the open source VOC2007 dataset and Ex Dark dataset,divide the training set and test set according to the 9:1 calibration,and use the depth separable convolution MBConv to construct the feature extraction network,and use the Bi FPN network to add the attention mechanism Network SE performs in-depth feature extraction,uses Momentum-Ada Grad as the optimizer of the regression frame,uses AP,m AP,and Recall as evaluation standards,and tests the detection accuracy of 20 different obstacles such as people,motorcycles,and cars.And compared with the traditional network,the final detection result m AP is 0.6983.In the second half of the experiment,the road video test was carried out on the model,and the experimental test was carried out on various road scenes collected.The results show that the parameters of the method proposed in this paper are good,can effectively reduce the training time,and have reliability in the detection accuracy of obstacles in road scenes.