SSD Pedestrian Target Detection System Optimized By Knowledge Distillation

Pedestrian detection is one of the hotspots in the field of computer vision.With the development of edge computing,deploying models on embedded platforms for real-time pedestrian detection has become a research hotspot.For the pedestrian detection algorithm on embedded platform,this dissertation selected SSD model for in-depth research and designed a pedestrian detection network suitable for embedded systems.First,in terms of network structure,in order to facilitate the implementation on the embedded platform,the backbone network of SSD model is replaced,convolution layer is added,and the lightweight SSD-M3 model is built,and the model storage is reduced by 4.52 times.At the same time,the knowledge distillation algorithm is introduced to construct loss function.The experimental results show that the accuracy of the KD-SSD-M3 model after distillation is improved,and the m AP value on the VOC07+12 and INRIA datasets reach 77.92% and 76.69% respectively.For 300×300 pixels input,the FPS of KD-SSD-M3 reaches 32.4 on NVIDIA Ge Force GTX 1660 Ti,which meets the requirement of real-time performance.Secondly,aiming at the problem of insufficient accuracy of SSD network.The SSD network is improved,and the channel-spatial attention module is added after the Conv4＿3,Conv7 and Conv8＿2 layers.At the same time,the feature fusion module is introduced to enhance the detection capability of the network.The experimental results show that the m AP values of the improved SSD model on VOC07+12,INRIA and Caltech datasets are 89.68%,91.12% and 78.35% respectively.The knowledge distillation algorithm is used to guide and train the students’ network,so that the m AP values of KD-SSD-M3 model can reach 82.16%,84.37% and 73.16%,which is basically the same as the original SSD network accuracy rate,and it can complete the pedestrian detection task in real-time and accurately.Finally,the pedestrian detection algorithm proposed in this study is transplanted to Raspberry Pi for verification.With Raspberry Pi as the main body,design the human-computer interaction software of the road pedestrian detection system,build Python and the corresponding environment,and use Py Qt5 to design a simple and easy-to-use operation interface.After the test,the pedestrian detection system based on Raspberry PI is accurate,and the FPS reaches 7.3,which basically meets the actual needs.