Research On Visual Object Detection Method For Autonomous Driving

In recent years,convolutional neural networks have brought completely new technical possibilities to autonomous driving.Unlike traditional logic algorithms,deep learning techniques can train end-to-end neural network models based on large amounts of data.This model has a wide range of applications,such as smart cars moving according to specific signs in warehouses,buses and private cars running normally on the road.A key point of autonomous driving systems is the perception of the environment.The correct recognition result is a necessary prerequisite for the safe driving of cars.This paper uses two types of data,image and point cloud,to study the visual object recognition method for autonomous driving,in order to improve the recognition accuracy and inference speed of the analyzed object.The analysis objects mainly include pedestrians,roads,and cars.The main contents of the paper are as follows:(1)For pedestrian recognition,this paper proposes an encoding method based on adaptive bin space to effectively deal with different objects or sparse data distributions.This paper introduces Siamese network to the data fusion process,which improves the network’s data prediction capability in the second phase.(2)For road recognition,this paper proposes TC convolution module and PC adjustment module to effectively and quickly extract global and local features.This paper proposes a lightweight backbone network with fewer parameters and fast computation speed.(3)For car recognition,this paper proposes a full view 3D backbone network.On the basis of analyzing 3D voxel data,this paper combines the features of several 2D views to comprehensively extract the features of point cloud data.The method proposed in the paper was tested on a public dataset,and the experimental results verified the effectiveness of the method.In order to facilitate the application of the proposed method in practice,a visual object recognition verification system is designed in this paper.The output prediction results are helpful for the automatic driving system to make reasonable decisions.