Research On Pedestrian Detection And Its Refined Attribute Recognition

Pedestrian detection has always attracted the attention of researchers due to it plays an important role in many practical applications such as automatic driving,intelligent surveillance and intelligent robots.Recent years,with the improvement of computing power and the expansion of annotated datasets,many pedestrian detection algorithms have made good progress by applying deep learning technology.However,in practical application scenarios,pedestrian detection still faces difficulties like multi-scale,pose variation,dense and so on,which will affect the detection accuracy.In addition,the further development of computer vision technology makes it possible to extract high-level semantic information from the detected pedestrians,that's pedestrian attribute recognition.Research on it can greatly promote the construction of intelligent security.Although lots of works has been proposed on this topic,however,pedestrian attributes recognition is still an unsolved problem due to huge challenging factors in real scenes,such as multiviews,low resolution,blur and so on.In order to solve the above problems,this thesis deeply studies two aspects,namely,the pedestrian detection algorithm based on convolutional neural network and pedestrian attribute recognition algorithm.In pedestrian detection,the improved algorithm based on YOLOv3 general framework has the following characteristics:1.Conforms to the prior of pedestrian morphology and size distribution,which is advantageous for the algorithm to converge to a better accuracy.2.Training with the integrated comprehensive dataset,which is beneficial to improve the quality of features learnt by network.3.Implement soft non-maximum suppression by introducing confidence penalty based on IOU,and apply it to the post-processing flow of the detection,which enhances the algorithm's detection ability in dense pedestrian scenes.The improved pedestrian detection algorithm in this thesis has achieved a miss rate of 10% on the public test set Caltech and a processing speed of 45 FPS on the 1080 Ti.As for pedestrian attribute recognition,this thesis mainly does the following work:1.Design and implement a pedestrian attribute recognition algorithm based on multi-label classification.Besides,aiming at solving the challenging factor of unbalanced data distribution,use the weighted cross-entropy loss function to improve the loss of multi-label classification.Evaluated on RAP dataset,the proposed algorithm has achieved a mA of 74.06,an accuracy of 65.44 and F1 value of 78.02.2.Solve pedestrian fine recognition problem by means of part detection.Propose two schemes of pedestrian part detection and recognition,namely,stagedscheme and joint scheme.And the implemented two-stage algorithm has achieved a test AP of 39.44 with the COCO evaluation standard.