Research On Pedestrian Detection Algorithm For Small Target Using Deep Learning

Pedestrian detection technology is an important branch of computer vision.It is defined as a computer determines whether it contains pedestrians for a given video or image.If it contains pedestrians,it needs to give the specific location of the pedestrians in the video or image.Pedestrian detection technology has a wide range of application scenarios in the field of video surveillance,autonomous driving and robotics.Small target pedestrians are the pedestrians whose height is relatively small compared to the height of the video or image.The resolution of small target is very low,its target feature that can be extracted is limited,and it is more susceptible to noise leakage,which resulting in miss detection.In addition,the small target requires greater search depth and the detection speed is affected.The difficulty of detecting small target impedes the application of pedestrian detection.Recently deep learning method has become the mainstream method in the field of pedestrian detection technology.Therefore,researching pedestrian detection algorithm for small target based on deep learning has important theoretical significance and practical application value.This paper is mainly to solve the problems of SSD algorithm in small target pedestrian detection.The main research work is as follows:1?A two-level feature fusion algorithm for enhancing the small target' detection performance of SSD is proposed.The algorithm adds a two-level feature fusion module to SSD algorithm,it makes full use of the context information between the feature layers,and it effectively integrates the deep network feature information into the shallow feature.The experimental results on Caltech pedestrian dataset and Campus pedestrian dataset show that the fusion strategy proposed by this algorithm is better than the fusion strategy of other algorithms.2?A small target pedestrian detection algorithm based on SSD-ResNeXt50 is proposed.The algorithm enhances the feature expression of the shallow network and reduces the miss rate for small target pedestrians by replacing the basic network of the SSD algorithm with ResNeXt50,adding the fusion module and the prediction module.The experimental results on Caltech pedestrian dataset and Campus pedestrian dataset show that the proposed algorithm can effectively improve the detection performance of small target pedestrians.3?A real-time and lightweight pedestrian detection algorithm for small target is proposed.The algorithm through replacing the basic network,re-selecting the feature extraction layer and setting the default box,and improving the loss function to reduce the model parameter and computational complexity of the SSD algorithm and enhance the detection performance for small target pedestrians.The experimental results show that the algorithm has better detection performance for small target pedestrians,whether on Caltech pedestrian dataset or Campus pedestrian dataset,and the parameter of the model is fewer,the training memory is smaller,the detection speed of the algorithm is faster.