Research On Indoor Personnel Detection And Tracking Algorithm Based On RGC Mask R-CNN

Nowadays,the living standards of our people have been greatly improved.At the same time,people's awareness of security is also constantly strengthening,especially in large indoor public areas,people want to be able to detect and track the target personnel in the indoor conditions of the video in real time to further make emergency judgment and processing.The rapid development of computer vision technology has made indoor personnel detection and tracking technology widely used in human-computer interaction,security monitoring and other fields.In the complex reality background,indoor personnel detection and tracking technology will inevitably be affected by personnel occluded completely or partially,personnel posture and so on.How to detect indoor target personnel accurately and achieve rapid personnel tracking is the starting point of this article.This paper summarizes some basic algorithms for indoor personnel detection and tracking,introduces their basic principles,and then proposes the improved algorithm,and makes relevant experiments and analyzes the results.The work of this paper is mainly summarized into the following two aspects.(1)An indoor scene personnel detection method based on RGC Mask R-CNN(residual network-group normalization-cascade instance segmentation)is proposed.The size and appearance of the indoor personnel are subject to change,and there is a problem that the person is blocked by the object and the person.This paper proposes the RGC Mask R-CNN model to detect people in indoor scenes.Firstly,the group normalization is added to the backbone network of the instance segmentation model Mask R-CNN,the side connection of the feature pyramid network,and the functional network of the bounding box.This paper follows the alignment of the regions of interest that play a key role in Mask R-CNN,and then R-CNN is set to cascade training during model training.Three different intersection-over-union thresholds are used to perform experiments on indoor multi-camera pedestrian data sets.The results show that the RGC Mask R-CNN is superior to the traditional instance segmentation method,and the bounding box accuracy of the model reaches 0.455,and the mask accuracy reaches 0.392.(2)An indoor personnel tracking strategy combining face recognition tracking and particle filter tracking is proposed.Face recognition tracking is useful for situations where a human face can be captured.When the face information cannot be obtained,this paper proposes an indoor scene personnel tracking method based on spatiotemporal context combined with particle filtering.Firstly,a spatiotemporal context method based on RGC Mask R-CNN feature extraction is proposed.Then,a block target model based on multi-cue fusion is established.The implementation of the target block greatly reduces the dimension of the network input,thereby reducing the computational complexity of the network training;in the tracking process,the model can dynamically adjust the weight according to the confidence of each sub-block,and improve the adaptability of the complex situation of the person's posture change and occlusion.Experiments on the OTB-50,OTB-100 and VOT2016 datasets show that the proposed algorithm outperforms other enumeration methods with an average overlap rate of 0.78 and more accurate indoor tracking.