Target Recognition In Non-Line-of-Sight Scenarios

Non-line-of-sight imaging is an emerging direction in the field of current laser imaging.Through this technology,it is possible to image the area outside the field of view of the optical system,which is of great significance for enhancing anti-terrorism capabilities and autonomous driving capabilities.In recent years,with the improvement of the accuracy of detection instruments and the advancement of image reconstruction algorithms,there has been an increasing tendency to deploy non-line-of-sight imaging to practical applications and a rising demand for target recognition in non-line-of-sight scenes.However,there are still two problems with non-line-of-sight recognition.On the one hand,the existing image restoration algorithms cannot better balance the real-time performance and image restoration quality of the algorithm.On the other hand,the nonline-of-sight restoration images are generally of poor quality.The shape distortion of the object is serious,making it difficult to identify.Therefore,this dissertation studies how to reconstruct target images suitable for recognition and how to recognize non-line-of-sight images.This dissertation studies how target images could be reconstructed to suit.First of all,this dissertation constructed a corresponding laser propagation model for non-line-of-sight experimental scenarios.Based on this theoretical model,the experimental scenarios are simulated,so as to simulate objects in different positions and different categories in space to generate the detector data.At the same time,it is considered that the existing image restoration algorithms cannot better reconstruct the reconstruction speed and accuracy.Therefore,by introducing the Meanshift positioning algorithm,this dissertation combines the virtual wave and the ellipsoid back projection to obtain a higher quality point cloud image at a faster speed,and then generates a series of training samples and test sample images.This provides sample materials for the feature extraction of non-line-of-sight images.Secondly,this dissertation analyzes the target characteristics of the non-view field.The corresponding filtering and projection operations are performed on the point cloud imagine since the image quality obtained by non-view field imaging is generally poor.By doing so,the noise in the imagine could be effectively reduced.Then extract the target information from multiple angles of shape feature,texture feature and depth feature,which helps to develop a more comprehensive description of the characteristics of the object and utilize the given information to an optimal extent.Lastly,use the mutual information evaluation index to quantitatively explain reliability of selected features.Finally,this dissertation introduces support vector machine model to identify nonline-of-sight targets.The particle swarm optimization algorithm will be introduced as the traditional support vector machine could be easily influenced by on the selection of artificial parameters,which is not conducive to practical application.In this way,the parameters could be intelligently selected and the level of human intervention will be greatly reduced.After considering that the classification model that favors a single feature is greatly disturbed by the change of the sample set,this dissertation analyzes the target characteristics under non-line-of-sight scenes and selects the Stacking strategy to integrate the classifiers composed of different features,thereby greatly enhanced the robustness of the algorithm.Finally,the dissertation will quantitatively illustrate the reliability of the recognition algorithm by introducing ROC and other evaluation indicators.