Research On Pedestrian Detection Technology In Vehicle Video Based On Deep Learning

Pedestrian detection technology in car video is an important part of driverless driving,and has high requirements in terms of detection accuracy and speed.Due to the scale problems of pedestrians in video or images and their background environment,the performance of the detector is greatly affected.This paper improves on the basis of the existing SSD algorithm to improve the performance of the detection system.The main research work includes:First,build AMSSD(Advanced MobileNet SSD,AMSSD)pedestrian detection network model based on SSD algorithm.Firstly,the MobileNet network is used in the SSD algorithm to reduce the amount of computation,thereby improving the detection speed.Then,the appropriate network layer is extracted to form the branch detection network,and the feature maps of different network layers are combined for the detection of the pedestrian scale.In this paper,the VOC and COCO datasets are used as training sets to test the Cityscapes dataset.The experimental results show that the AMSSD algorithm can improve the detection speed while maintaining accuracy,and achieve real-time effects.Secondly,based on the AMSSD pedestrian detection network model,the Focal Loss function is added to construct the FAMSSD(Focal Loss and Advanced MobileNet SSD,FAMSSD)pedestrian detection network model.This model balances the number of positive and negative samples and simple difficult samples in the algorithm for better pedestrian detection performance.The experimental results show that the accuracy of the FAMSSD pedestrian detection algorithm is better than that without the Focal Loss function.