Research On Video Object Detection Algorithm Based On Lightweight Deep Learning

With the continuous construction of safe city and smart city in China,massive video data was brought because of the extensive use of surveillance cameras.It's difficult for traditional manual monitoring to acquire effective information from video in time.Video analysis technology based on video object detection plays an increasingly important role.With the more convenient collection for data and a significant increase for computer compute capability,deep learning has rapidly become the mainstream research method of computer vision with the excellent representation and generalization capabilities.Compared with image,video has the characteristics of massive amount of data,sequential correlation of adjacent frames.Video object detection also faces challenges,such as motion blur,video out-of-focus,and partial occlusion.In the paper,the network structure of object detection in video is improved to reduce the inference delay,and the temporal characteristics of video is used to improve the quality of object detection in video.The balance of accuracy and speed is achieved.The main work is as follows:(1)MEfficient Det is proposed to improve the lightweight object detection model Efficient Det-D0 for the detection speed.The detection speed of MEfficient Det is increased while losing a little performance.MEfficient Det uses the activation functions Leaky Re LU in the shallow network and Hard-Swish in the deep network,which cost less time,and removes the image attention SE module that has a small impact on performance but contains multiple 1x1 convolution operations,and introduces A cross-stage partial connection structure is used to enhance the learning ability of the lightweight network.Through the comparison experiments with the typical algorithms of object detection,like Faster R-CNN and YOLOv4,and the ablation experiments with the Shuffle Net as the lightweight backbone network,MEfficient Det loses a little accuracy but significantly reduces the memory usage and inference delay compared with base model Efficeitn Det-D0.The effectiveness of the MEfficient Det is verified.(2)Combining the temporal characteristics of video,an online object detection model in video MEfficient VDet and a post-processing method of the result feedback are proposed.The current frame by fusing temporal feature with the previous frame using the adjacent frames' correlation coefficient as fusion weights,which learned by MEfficient VDet in different circumstances,is improved the problem of insufficient feature information of video object that is difficult to be detected due to factors,like video motion,out of focus,and occlusion.The post-processing of the result feedback is proposed to improve the missed object problem of the low-confidence object caused by the fixed threshold,by calculating the Io U value of the detection result of the previous frame and the candidate results of the current frame,and rescoring the confidence of objects of which Io U exceed the threshold.The quality of video detection could be improved through these two methods.These methods only rely on the pre-sequence frame results,and could process the video stream online without waiting for the complete input of video clips.Through comparison experiments with classical object detection in video,such as DFF,FGFA,and Seq-NMS,the effectiveness of the methods is verified.(3)A simple object detection system for surveillance is designed and implemented.A simple and intuitive Graphical User Interface and functions of tradictional surveillance system,such as playback,pause,and adjustment of play speed are provided.The method(2)is also used to detect the object of interest in video data.The frame sample by differential similarity is used to avoid redundant calculations for static and slow object.The functions and performance of the system were tested by using local traffic video and streams from camera,which verified the effectiveness and practicability of the system.