| The continuous development of electronic technology has led to the shrinking size and increasing performance of mobile devices,which enables many performance-demanding projects to be implemented on mobile devices.The application prospect and market value of deep learningbased target detection and target ranging technology in artificial intelligence and other fields are growing day by day,and the application in various industries is becoming more and more extensive.Based on the large number of visually impaired people in China,it is difficult for visually impaired people to move around in indoor environment,based on this,this paper conducts in-depth research on the key technologies in target detection and target ranging to build an indoor scene recognition system,the main research work is as follows:(1)In the target detection part,Center Net is used as the benchmark algorithm to improve the existing target detection algorithm for the problem that it relies on a large amount of computing resources when running.The original backbone network is replaced with Efficient Net V2-S to realize the light weight of the backbone network and improve the speed of feature extraction from the backbone network,while the convolutional attention module is introduced into Efficient Net V2-S in order to enhance the feature extraction capability of the Efficient Net V2-S network;for the indoor scenes where the same targets are easily stacked and In order to improve the detection accuracy,Soft-NMS is introduced to post-process the detection results;the loss function is optimized and variable parameters are introduced to adjust the output of the loss function to reduce the prediction error caused by the imbalance of positive and negative samples in indoor scenes.The experimental results show that compared with other classical algorithms and lightweight algorithms,the running speed of the algorithm in this paper has obvious advantages.Compared with the original algorithm,the running speed of the improved algorithm is increased by 27.24%,the model size is reduce significantly,and the detection accuracy is basically the same as the original algorithm.(2)In the target ranging algorithm section,the key technique of stereo matching algorithm is optimized.Matlab software is used to calibrate the binocular camera and obtain the parameters of the binocular camera to eliminate the influence of the camera’s own aberrations on the ranging results.The fourth stage post-processing module is replaced by a faster convolutional space propagation network to balance the running speed of the improved algorithm.optimization.The experimental results show that,compared with the original algorithm,the error of the improved algorithm decreases at each stage of calculating parallax maps with different resolutions and runs faster,and in the fourth stage with the highest accuracy,the endpoint error of the proposed algorithm decreases significantly and the running speed increases by 26.67%,compared with other similar In the fourth stage with the highest accuracy,the endpoint error of the proposed algorithm decreases significantly and the running speed increases by 26.67%.(3)In the system building part,the Nvidia Jetson Xavier NX developer kit is selected as the deployment device for the algorithm,and the target detection algorithm and target ranging algorithm are ported and tested.The target detection algorithm and the target ranging algorithm were tested separately,and the experimental results showed that both the target detection algorithm and the target ranging algorithm could run stably.The fusion algorithm of target detection and target ranging was tested in different scenes,and the experimental results showed that it can accurately broadcast the name and distance of detected objects in indoor scenes,which can be used to facilitate the life of visually impaired people in indoor scenes. |
