Research On Multi-object Tracking And Segmentation Algorithm Based On Depth Video Flows

With the improvement of theory and computer capability,substantial achievements have been got.Especially in the field of instance segmentation and multi-object tracking,significant breakthroughs have been made.In recent years,more and more attentions have been paid to Multi-Object Tracking and Segmentation in a neural network simultaneously.The research in intelligent breeding,automatic driving and intelligent medical treatment has wider application prospect.Especially in intelligent breeding,Multi-Object Tracking and Segmentation algorithm could help to improve the accuracy of non-contact weight and backfat measurement.This paper introduces the research status of Multi-Object Tracking and Segmentation,related algorithms and theories.And research from the following aspects:First,improve a Multi-Object Tracking and Segmentation network Track R-CNN which based on two-stage algorithm.To solve the problem that the edge of the predicted mask is straight that do not fit outline of targets well.Mask generation branch of the network is redesigned by combining the atrous convolution and deconvolution flexibly,then its prediction masks are more suitable for the targets.For ID switching when targets are close,feature maps reweighting module is added in association embedding branch.Useful features are enhanced and useless features are suppressed.More accurate association vectors are got so that tracking errors could reduce.Thus,a high-precision Multi-Object Tracking and Segmentation network was accomplished.Then,design a high-speed Multi-Object Tracking and Segmentation network based on one-stage instance segmentation network YOLACT++.Appearance features extract from instance segmentation network,together with bounding box coordinates and class to do cascade matching and IOU matching to get tracking ID.Meanwhile,in order to predict more accurate prediction masks,the prototype mask network was redesigned by global attention upsampling module.Combining the prediction mask and tracking ID could get the final Multi-Object Tracking and Segmentation results.The whole network could do online Multi-Object Tracking and Segmentation in real-time.Finally,make dataset for Multi-Object Tracking and Segmentation with group housing pigs' data,train and test model.Visualize results and calculate s MOTSA,MOTSA and MOTSP.Compare the speed and accuracy of the above algorithms in depth video flows data which lacks texture information.