Research On Pedestrian Detection Based On Automatic Driving Scene

With the development of artificial intelligence and 5G technology,autonomous driving has gradually entered the public’s attention after years of technological precipitation.Autonomous driving systems can alleviate road traffic congestion,improve road safety,and provide more travel options for people with limited mobility.The main protection object of the autonomous driving scene is people,including drivers and passengers,but also pedestrians on the road.The system needs to quickly and accurately detect persons and estimate the distance,taking emergency response accordingly,such as whistle warning or braking.Therefore,pedestrian detection is of great significance in autonomous driving.Over the last few years,deep learning has developed rapidly.It has been widely applied in various visual tasks.Pedestrian detection,as a special target detection task,has also received extensive attention recently.The main difficulties of pedestrian detection in autonomous driving scenarios are occlusions,inconsistent scales,and changeable environments.This paper proposes a pedestrian detection framework without anchors,which uses key point positioning to detect two corner points and their connection to determine the final location of pedestrians.First,the structure of the feature pyramid is designed according to the multi-layer output of the basic network.The hybrid features containing multi-scale information and context information are obtained by fusion and reorganization of deep features and shallow features,and then a dense hybrid dilated convolution module is constructed to increase the receptive field.Pyramid has a strong ability to express pedestrian characteristics of different scales.This paper also proposes a learnable spatial attention module to generate the spatial attention map through simple convolution operations.The attention map is used to increase the weight of the target area and weaken the weight of the background area,which can enhance the perception of occluded targets.Comparative experiments are conducted on several public pedestrian detection datasets which are commonly used in autonomous driving scenarios,so as to demonstrate the availability of the proposed network.The results illustrate that the problems of scale variations and occlusions can be efficaciously solved and the proposed algorithm is superior to existing pedestrian detection methods in indicators such as missing rate and speed.