Research On Fast Pedestrian Detection Algorithm

Pedestrian detection is the use of computer vision technology to determine pedestrians in an image or video and to mark the precise location of the pedestrians.With the rapid development of deep learning,deep convolutional neural network algorithms have become the mainstream direction for researchers to solve pedestrian detection problems.Because pedestrians have both the rigid characteristics of the object and the flexible characteristics,the appearance is easily affected by posture,occlusion,wearing,and vision.Pedestrian detection is a challenging task.In a real-life scenario,only when the real-time detection speed and high detection performance are satisfied at the same time,the abnormality can be found and the warning can be performed in time.To this end,it is a practical work to develop a faster and better performance pedestrian detection algorithm.In this thesis,the target detection algorithm RefineDet which performs well in both detection performance and detection speed is applied to pedestrian detection as the research object,and the network structure is improved.A multi-feature maximum fusion method of RefineDet is proposed,referred to as MFMF-RefineDet.The MFMF-RefineDet algorithm proposed in this thesis is verified by the pedestrian detection dataset Caltech.For the input image of 320?320,the Miss Rate is 16.9%,which is 7.7% higher than the original RefineDet algorithm,and the detection speed reaches 0.026 seconds per image;for the input image of 512?512,the Miss Rate is 12.7%,which is 2.6% higher than the original RefineDet algorithm,and the detection speed reaches 0.057 seconds per image.Compared with some well-known pedestrian detection algorithms,the MFMF-RefineDet algorithm proposed in this thesis greatly improves the detection speed,and the proposed MFMF-RefineDet algorithm has a higher detection performance in the case where pedestrians are smaller and severe occlusion.In order to reduce the computational complexity of convolutional neural networks,this thesis applies lightweight networks to pedestrian detection.This thesis proposes four pedestrian detection algorithms for lightweight networks: MobileNet V1-RefineDet,MobileNet V2-RefineDet,ShuffleNet V1-RefineDet,and ShuffleNet V2-RefineDet.This thesis mainly replaces the backbone network VGG-16 in the RefineDet algorithm with the lightweight networks MobileNet V1,MobileNet V2,ShuffleNet V1 and ShuffleNet V2,respectively.In order to maintain the consistency of the overall network structure,the network is adjusted.The experimental results show that MobileNet V2-RefineDet has the best detection performance and MobileNet V1-RefineDet has the fastest detection speed among the pedestrian detection algorithms of four lightweight networks.For the MobileNet V2-RefineDet algorithm,this thesis adds an additional pre-training model,which improves the detection performance.Compared with the original RefineDet algorithm,MobileNet V2-RefineDet is very close in detection performance and detection speed,in terms of model size,MobileNet V2-RefineDet has a great advantage.