Object Detection Technology Based On Structured Light Depth Image

Object detection technology is widely used in public safety,military defense,medicine and other fields.This paper focuses on the problem that low accuracy of object detection using a single structured light depth image or color image while using the original low-resolution structured light depth image and color image.The following research work was carried out on super-resolution reconstruction of images,object detection based on structured light depth image and color image:(1)Aiming at the fact that SRCNN,ESPCN and other algorithms can not meet the requirements of better super-resolution reeonstruction and real-time processing at the same time,an new image super-resolution reconstruction algorithm based on convolutional neural network is proposed.The low-resolution image without any pre-processing is used as the input data,and the feature of the original image is extracted by the convolution operation,the 1×1 small convolution kernel is used to reduce and enclose the feature map to reduce the parameters of the network and improve the reconstruction effect;the combination of deconvolution and pooling is used to zoom in and out the feature map to extract features that are more sensitive to the results;finally,the reconstruction is performed by deconvolution.Experiments show that the proposed algorithm not only achieves better reconstruction results,but also processes more than 24 images with a size of 320 X 240 per second,which meets the real-time requirements for video super-resolution reconstruction.(2)Aiming at the problem that the SRCNN algorithm has weak learning ability and poor reconstruction performance and slow network convergence,an image super-resolution reconstruction algorithm based on deep residual network is proposed.A deep network consisting of 42 convolutional layers is designed.The combination of the Inception structure and the residual network is used to leam the residual at the upper layer of the network,and then,the residual is added to the input data to get the final output.The Inception structure increased the network width,thus increased the nonlinearity of the network and improved the reconstruction performance,and the residual network accelerated the convergence speed of the network.Experiments show that compared with the Bicubic and SRCNN,the proposed algorithm achieves better results with fewer training iterations,and the network convergence speed is significantly improved.(3)Aiming at the problem that the object detection method using color image or structured light depth image alone has some limitations,resulting in unsatisfactory detection performance,a two-stream SSD detection network combining structured light depth image and color image is proposed.Based orn the original SSD(Single Shot MultiBox Detector)detection network,the depth map channel network is added,and the depth image and color image are merged in different ways in different network layers.The experimental results show that the fusion of depth image and color image is beneficial to object detection,making full use of color information and depth information,which can improve the accuracy of object detection.(4)Aiming at the problem that the object detection method using the original structured light depth image and the color image is affected by the low resolution of the image,the detection accuracy is not high,based on the two-stream SSD detection network proposed in this paper,a detection network combining high resolution depth images and color images is proposed.Firstly,an algorithm with better reconstruction effect is selected from the super-resolution reconstruction algorithm proposed in this paper to reconstruct the original depth image and color image,and then,the reconstructed depth image and color image are combined for object detection.Experiments show that the reconstructed structured light depth image and color image can further improve the accuracy of object detection.The two reconstruction methods proposed in this paper can improve the performance,and realized real-time processing and acceleration convergence respectively;the proposed two-stream SSD detection network improved the detection accuracy;the original image is processed by the better method in the reconstruction method proposed in this paper,and the obtained image is combined to perform object detection,which further improved the detection accuracy.