Research On Deep Learning For Pedestrian Detection

Pedestrian detection refers to the accurate identificatiorn of pedestrians in pictures or videos through algorithms,which is an example of target detection.Because it has important applications in robotics,driverless,intelligent security and intelligent monitoring,it has become a key research.Although it has greatly improved in the detection performance after years of development,in actual scenes,lots of facts including pedestrian's clothes,light,background,their body posture,and occlusion can cause their appearance to change greatly,which makes the recognition difficult.The main steps of the traditional pedestrian detection method are to preprocess the image first,then extract features,select features,and finally train the classifier to predict.The disadvantage is that the region selection strategy based on the sliding window is not targeted,the time complexity is high,the window is redundant,and there is no robustness to the change of the target appearance and scene.Traditional methods usually combine multiple low-level features without using higher-level features of the image,and their detection accuracy and speed cannot meet real-time requirements.With the great success of deep learning in the field of target detection,there is a large amount of literature recently that shows that the use of deep learning as a pedestrian detection can improve detection performance.This article has carried on further research on the existing basis.In order to fix the problem of poor detection performance for small targets which exists in a classical framework for target detection named Fater-rcnn,this paper proposes a deep learning pedestrian detection method based on hybrid multi-scale network.This method consists of two parts,one is the target region extraction,the other is target detection.In the network structure of target extraction and target detection,a multi-scale method was adopted.By comparing single-scale and multi-scale,multi-scale method was found to can improve detection accuracy.In the experiment,the effect of different detection network sizes on performance was compared..This method has a missed detection rate of 10.61%on the Caltech dataset,an undetected rate of 35.14%on the ETH dataset,and an undetected rate of 10.48%on the INRIA dataset.The depth of the neural network has a great influence on the performance of the target detection.Usually increasing the depth of the network can improve the performance of the neural network.This paper proposes a deep learning pedestrian detection method based on channel selection.In this method,the SE module is added to the VGG16 network of the SSD target detection framework.Instead of using traditional convolutional results of the input channel,a simple superposition is performed,and the correlation between the channels is sought to give a different channel.The weights are then weighted and summed.The pedestrian detection system composed of this method achieves good performance.Experimental results show that the method has a missed detection rate of 10.01%on the Caltech dataset,a missed detection rate of 32.08%on the ETH dataset,and a missed detection rate of 9.28%on the INRIA dataset.