Research On Pedestrian Detection And Density Estimation For Low Altitude Monitoring Image And Video

In recent years,machine learning has been into the boom because of the quick development of hardware and the improvement of the algorithms.In the meantime,computer vision is always a popular research direction and many research outcomes have been utilized in various aspects of our lives.Machine learning combined with computer vision can provide many solutions for practical problems.Nowadays,the most traditional way to capture video or pictures is by using a camera fixed at some place.During the past few years the low altitude flight platforms,such as unmanned aerial vehicle,have been developed very quickly so that much more videos and pictures can be captured through those tools.Thus,the pedestrian detection on those image and video can be a meaningful subject.In the above background,this dissertation mainly researches on pedestrian detection on the image and video captured by low-altitude flight platforms.As for the pedestrian detection based on the video,this dissertation proposes a method which combines HOG+SVM algorithm and low-rank matrix decomposition algorithm.The new method can improve the original HOG+SVM algorithm.The sparse part of the continuous frames of a video can be extracted through low-rank matrix decomposition.After some kinds of processing,those parts can be used in the HOG+SVM detection.The experiment result shows that the new method can cut down the error detection ratio while maintain the recall ratio.As for the pedestrian detection based on the image,this dissertation modifies the frame of the original Fast YOLO.The original Fast YOLO uses darnet interface and this dissertation changes it to Caffe interface.Besides,the original Fast YOLO has just nine convolutional layers,thus the feature extraction is not overall.This dissertation refers to GoogLe Net frame and modifies it to enhance the width and depth of the network.The experiment result shows that the modified network can improve the recall ratio while cut down the error detection ratio.This dissertation also designs and implements a pedestrian detection system.The system consists of five modules,including real-time preview,cradle head control module,pedestrian detection module,POI location module and map location module.The system can capture pictures and videos,and perform pedestrian detection function.Besides,the system has some other functions.