Design Of Head Detection Algorithm Based On Scale Estimation

At present,the detection of human head has been widely used,such as the traffic statistics of shopping malls,the statistics of student classroom attendance,the statistics of subway station and train station traffic.At the same time,human head detection is an important part of the tasks of pedestrian detection and pedestrian recognition.At present,most of the head detection is based on the traditional target detection method,the feature extraction ability is weak,and the sliding window structure is needed.The detection rate and accuracy are generally low,the generalization ability is weak,and the time is serious.Aiming at the problems existing in the traditional human head detection algorithm,this paper proposes a human head detection algorithm based on deep learning.The main research contents include:Overall algorithm framework design.This paper uses Caff to build the main algorithm framework.Firstly,the scale estimation network is designed to predict the size of the head size in the input picture,and to avoid the calculation of unrelated size images in the subsequent network.Then the full convolutional network(FCN)structure is designed to classify and regress the whole picture of any size,which avoids the extraction process of the sliding area to the image area,which greatly reduces the time-consuming of the algorithm.Finally,two cascaded convolutional neural network structures are designed to fine-tune the output of the FCN output.Between the modules,a mapping algorithm that maintains the shape of the feature shape and a non-maximum value suppression(NMS)algorithm that eliminates the redundant classification box are also designed.Algorithm optimization.Firstly,a data cleaning method based on K-means clustering is proposed for the missing data.The accuracy and detection rate of data cleaning are increased by 1.5%.Then,using t-SNE to visualize the dime nsionality reduction of the data,by analyzing the dimensionality reduction of the two-dimensional data distribution map,it is found that the misdetection of the algorithm is mainly caused by the human head similarity between the positive and negative samples.Through the above work,a head detector with an accuracy rate and a detection rate of 91.5% and a single frame time of 60 milliseconds is finally obtained.