Research On Object Detection Method By CNN Based On Temporal-Spatial Information

In recent years,the research on deep learning has made breakthroughs.The object detection task based on deep learning has a wide range of applications in computer vision,such as vehicle imaging,robot navigation,driverless,etc,so object detection has become a theoretical study and Application hotspots.At present,the single-frame based object detection method has achieved good results,and some methods can perform real-time detection.However,there are still some problems in directly using this method in video-based object detection,such as blurring and out of focus in video frames,etc,the object detection method based on a single frame cannot detect such an object.Therefore,the object detection method is studied in this paper.By using the correlation characteristics between adjacent frames in the video,the temporal information and spatial information are combined,and the convolutional neural network model is improved to improve the accuracy of object detection.In object detection based on convolution neural network,the YOLOv3 method of convolution neural network is mainly improved.In this paper,through the research and analysis of the YOLOv3 method,in order to solve the problem that the YOLOv3 method is not ideal for detecting medium or large scale objects,it is prone to missed detection and false detection.This paper proposes an improved YOLOv3 method after the main network.Two residual blocks are added,and the characteristic pyramids with five different scale convolutional layers are constructed together with the original three different scales of YOLOv3,and predictions are made on the five different scale feature maps.Through method improvement,the accuracy of detection of medium-scale or large-scale objects is improved while detecting small objects.In terms of object detection in video,the video contains rich temporal and spatial information,and objects based on single-frame images.This information is not considered in the detection,and it is impossible to detect a blurred or out of focus object in the video frame image.Therefore,in order to improve the accuracy of object detection,this paper proposes a temporal-spatial information fusion method.In this paper,the video frame image is preprocessed by temporal and spatial information respectively.The optical flow method is used in time,and the method based on RC method is used in space,and then the method of weighted fusion using temporal-spatial information will be used.temporal-spatial information is fused.Finally,using the improved convolutional neural network model for object detection,the detection results are obtained.