Application Research Of Object Detection Based On Deep Learning

In recent years,the object detection algorithm model based on deep learning has been continuously improved in network accuracy and computing speed.However,in some specific scenes,only relying on the image information obtained by the color camera and the two-dimensional object detection algorithm can not meet the needs of practical applications.In industrial production and life,the detection of colorless,odorless,irregular air and liquid leakage is very important,but the color camera is difficult to capture the characteristic information,so the detection algorithm based on infrared image is particularly necessary.With the continuous development of robot technology,scientists and technicians begin to use mobile robots for space exploration.The terrain perception of the robot’s surrounding environment can help the robot identify the object and obtain the surrounding information.The spatial positioning of the object of interest can help it carry out the next exploration.At the same time,in the field of automatic driving,the two-dimensional object detection algorithm is limited to two-dimensional features,while the three-dimensional object detection algorithm based on deep learning can directly extract features from the point cloud information and obtain the spatial information of pedestrian,vehicle and other objects in the scene.For the above three practical application scenarios,this paper puts forward objected solutions around the deep learning object detection method.The specific research contents are as follows:(1)Aiming at the problem of gas and liquid leakage detection,a gas and liquid leakage detection method based on infrared image is proposed in this paper.Firstly,the infrared information of the object is captured by the infrared thermal imager,and the infrared image data is collected by the infrared imaging system.Aiming at the characteristics of low signal-to-noise ratio and unclear features of infrared image,the opening and closing operation and image enhancement algorithm are used to suppress the background noise and enhance the detailed features of infrared image.The detection network uses the weight file pretrained by the MS coco data set as the initialization parameter,uses the clustering algorithm to obtain the prior knowledge of the anchor rectangle,and improves some feature extraction modules and feature fusion modules based on the YOLO v4 network,so as to improve the accuracy of the algorithm and meet the requirements of the system for the detection speed and accuracy of the model,Thus,the detection problem of air and liquid leakage object is solved.(2)Aiming at the problem of terrain segmentation and depth detection,this paper proposes a method of object location based on RGB-D spatial semantic data.Firstly,a terrain semantic segmentation algorithm based on cyclic consistency countermeasure network is proposed.Secondly,the terrain instance segmentation is realized by using mask r-cnn algorithm.Finally,the depth camera and color camera are projected into 3D space for spatial alignment,the point cloud information of the segmented specific object instance terrain is obtained,and its spatial position is calculated.At the same time,the down sampling and noise point cloud filtering method based on statistics are proposed to improve the accuracy of positioning.(3)Aiming at the problem of 3D object detection,this paper proposes an object detection algorithm based on image and point cloud data.Firstly,for the acquisition of point cloud data,in addition to the depth camera used in the previous chapter for projection,this paper also studies the binocular stereo matching algorithm based on depth learning,using binocular image to generate pseudo point cloud.Secondly,based on the two-dimensional object detection network,the threedimensional object detection algorithm based on the visual cone point cloud detection network is used to extract the semantic features of the point cloud corresponding to the detection frame to complete the three-dimensional object detection.To sum up,this paper mainly studies the specific application of object detection technology based on deep learning.Based on deep learning algorithm,relying on the actual application projects,using infrared sensor and binocular simultaneous interpreting data,and combining with deep learning algorithm,the corresponding solution is proposed.Experiments are conducted on selfmining dataset or open dataset to verify the effectiveness of the algorithm,and solve the application of object detection algorithm in related actual scenarios.