Research On Pedestrian Detection Method Combined With Semantic Features In Unmanned Driving

With the rapid development of artificial intelligence technology,pedestrian detection technology has attracted more and more researchers’ attention due to its wide application range and huge application value.The main task of pedestrian detection is to identify the location and scale of pedestrians in images or videos.This paper studies how to detect pedestrians in images with semantic features extracted by convolutional neural networks.The main research contents are as follows:We propose a pedestrian detection method(MF-CSP)that combines multi-layer features with center point and scale prediction.The model is mainly composed of semantic feature extraction module and detection head module.The semantic feature extraction module fully fuses shallow features and deep features in the feature extraction process,and further fuses feature maps suitable for detecting pedestrians of different scales,and obtains feature maps containing rich pedestrian information to be detected.The detection head module consists of three detection branches,two of which solve the pedestrian localization problem according to the semantic features of the pedestrian center point,and the other branch predicts the corresponding height.Experimental results show that this method has higher detection accuracy than other commonly used methods in different occlusion subsets of Caltech and City Persons pedestrian databases.Compared with the benchmark model,MF-CSP has a decrease in missed detection rates of 1.9% and 1.4% in the subset of pedestrians that are heavily occluded by Caltech and City Persons,respectively.We propose a pedestrian detection method combining attention mechanism and multiple semantic features.First,for the feature maps of different channel dimensions to focus on the characteristics of different regions of the human body,we design an attention module after extracting features from the backbone network to improve the model’s attention to the correct semantic regions.Secondly,in order to reduce the missed detection rate when the pedestrian center point is occluded,a pedestrian head prediction branch is designed in the detection head module,and the model’s confidence in the detection results of occluded pedestrians is improved by combining the two semantic features of head and center point.The results on the Caltech and City Persons datasets show that the method has good detection performance for occluded pedestrians and small-scale pedestrians.In the City Persons data,compared with the algorithm based on center point detection,the missed detection rate of this method in the subset of heavily occluded pedestrians decreases by 2.0%,and the missed detection rate of the small-scale pedestrian subset decreases by 1.2%.On the basis of the method in this article,we design and implement a set of pedestrian detection experiment system.The system provides two types of typical pedestrian detection algorithms,single-stage and two-stage.The main functions include uploading images or videos to be detected,setting detection hyperparameters,and detecting pedestrians in images.In order to verify the detection performance of the detection system in real scenes,we collect 1000 images from the actual scenes and produce a pedestrian data set.The results on the self-made database test set show that the joint attention mechanism and the multi-semantic feature pedestrian detection method proposed in this paper has good detection performance for pedestrians in actual scenes.In the heavily occluded and small-scale pedestrian subsets of the self-made database,the algorithm in this paper achieves the best missed detection rate.Compared with the two-stage Cascade R-CNN,the missed detection rate in the heavily occluded pedestrian subset decreases by0.59%,and the missed detection rate in the small-scale pedestrian subset decreases by 2.44%.