Home Object Detection Based On Anchor Redefinition And Feature Aggregation

With the development of artificial intelligence technology,service robots are gradually integrated into people's daily lives,but due to the complex family environment,robots are still hard to perform difficult family services.As the basis for the service robot to complete various tasks,object detection is a key link to improve its level of intelligence.Although the Object detection algorithm has excellent Performance,it still faces a series of difficulties when it is directly applied to the service robot object detection in the household environment.There are many kinds of household items with different shapes and sizes.It is difficult to manually set the prior box to accurately locate the object.The workload of dataset construction is heavy.The single frame detection effect is not great due to the motion blur,video defocus,and part occlusion during the image collection.The object detection algorithm based on deep learning requires high computing power.It is difficult to implement the algorithm Limited by the hardware conditions of the robot body.In response to the above problems,it is studied from the aspects of data set construction,network structure,video stream data application,algorithm deployment in this paper.Firstly,for the application of service robot object detection in the home environment,an object detection algorithm based on anchor redefinition is designed.Using CSPDarkNet as the backbone network,it can enhance the feature expression ability while reducing the computational pressure.To improve the detection effect of items of different scales,the path aggregation feature pyramid network is used.The prior box is obtained by k-means clustering instead of manual selection,and the prior box is adjusted by the anchor redefinition to enrich the size distribution of the prior box and meet the requirements of variable scales of objects.Secondly,in order to make full use of the temporal context information and redundancy of the video streaming,a video object detection method based on feature aggregation and motion estimation is proposed.The feature aggregation of adjacent frames is used to improve the sudden change in the detection effect caused by the bad quality of a single frame,and the key frame selection method is used to reduce the operation time of feature aggregation.The optical flow network is used to estimate the motion of the object,and the detection results is adjusted according to the local optical flow field nearby the object,and then propagated to the current frame to improve the recall rate.Target local and global alternate detection strategy based on the analysis of the object displacement data in the video stream is used to improve the detection efficiency.Thirdly,in response to the difficulty of constructing large-scale datasets,an automatic annotation method based on web crawler and background elimination is proposed,and multi methods is used to construct large-scale datasets.Rich instance images are obtained by web crawler technology and autonomous shooting platform.Household images are extracted from public scene datasets.According to the position of the mask obtained by the background elimination,images of dataset are composed and annotation files are generated automatically.A large-scale household object detection dataset is built by method that manual labeling is the main method with automatic generation and public dataset extraction as a supplement.Finally,in view of the limitation of robot hardware,a household object detection system based on private cloud is designed.For the image transmission,a wireless local area network is built through a router.The detection model runs on a high-performance computing platform.The robot is responsible for image collection.An asynchronous image transmission program are developed based on the ROS system.The system is tested in the real world.