Research On Peripheral Target Information Perception Of Nighttime Driverless Vehicles

The driverless vehicle environment awareness system is like a human sensory system,which is a source of information for driverless vehicles to make decisions.Machine vision is the most common method used by driverless vehicles for environmental awareness.Therefore,it is great significance to study the driverless vehicle environment awareness system based on visual image.At night,road visibility is relatively low,road information is difficult to obtain,and safety hazards are greater than the day.In the absence of light at night,ordinary visible light imaging device cannot obtain effective environmental information,and the imaging principle of infrared thermal camera is to obtain the temperature distribution of the scene,which has unique advantages for acquiring targets in nighttime scenes.The infrared image has less texture information and lower contrast,so it is difficult to study driverless vehicle environment perception in night scenes.Using the night machine vision to detect and identify the night target to obtain the information of the target around the driverless vehicle,it can expand the perception ability of the driverless vehicle at night,which effectively help the driverless vehicle to make corresponding decisions on the obstacle in time.This paper mainly studies the perception of target information driverless vehicles based on night vision scenes.The main research content includes two parts: the first part is the object detection and vehicle angle prediction based on the improved YOLOv3 network;the second part is research on depth estimation method of cascading convolutional neural network and multi-scale continuous CRF,combining object detection results with depth estimation to create the surrounding vehicle distance and velocity perception model.The main innovations of the paper are as follows:1.An improved YOLOv3 network capable of predicting the direction of vehicle driving is proposed to add peripheral vehicle angle information to the position information of the YOLOv3 network bounding box to form an end-to-end network.The network can realize the detection of pedestrians and vehicles in the infrared images captured by the driverless vehicle at night,which can significantly improve the speed and accuracy of object detection.The problem of estimation in the moving direction of surrounding vehicles is transformed into the problem of estimation of the angle of the surrounding vehicle position,and the driving intention of the surrounding vehicle is effectively predicted.2.A depth estimation algorithm based on cascading convolutional neural network and multiscale conditional random field is proposed.The algorithm exploits multi-scale estimations derived from CNN inner layers by fusing them within a CRF framework.The algorithm shows how mean field(MF)updates can be implemented as sequential deep models,enabling end-to-end training of the whole network.The realized depth estimation network is combined with the improved YOLOv3 network,and the target distance and speed information are obtained by combining the object detection results with the estimated depth information.