Research On Small Object Detection Algorithm Of Lidar Based On Deep Learning

Li DAR(Light Detection And Ranging)has a narrow and directional beam,which can pass through the barrier to obtain target information in a certain extent.Compared with traditional two-dimensional imaging sensors,Li DAR can obtain not only rich reflectivity intensity information of the target,but also 3d structure information and location information.In addition,GM-APD Li DAR has high detection sensitivity,greatly reduces the system volume and power consumption,makes the system have the feasibility of practical application.Small object detection for Li DAR images can help the system find and locate targets earlier,which has greatly application potential.At present,there are three difficulties in Li DAR’s object detection.Firstly,due to the limitation of the pixel number,the spatial resolution is low,and it is difficult to obtain the clear contour of the remote target,the object detection rate is not high.Secondly,object detection algorithms can only process a single type of images,and it is difficult to use the intensity information and the range information of the target at the same time.Finally,there is a high degree of similarity in the data set,the amount of the data is small,and there is no publicly data set.Training based on this data set has the risk of overfitting.Aiming at the above problems,the Li DAR data set are studied in this paper,Specific work and research are carried out from the following three aspects:Firstly,aiming at the problem that the amount of Li DAR images is small and the images have high degree of resemblance.The data augmentation is studied of the traditional trigger model-based method and generative adversarial network method,And a generative adversarial network combined with attention module is proposed.The intensity images and range images are input into the network as different channels to generate the samples.In the Li DAR data set,the algorithm achieves the best performance in both IS and FID evaluation indexes.The effectiveness of the data augmentation method is verified by the object detection network.In the Faster RCNN network,the detection accuracy is 88% before data augmentation and 90% after that,increasing by 2%.Secondly,aiming at the problem that the detection accuracy of intensity images is not high.This paper analyzes the reasons of the performance and proposes a small object detection algorithm based on the feature pyramid network.In this algorithm,attention module and receptive field module are added into the feature pyramid network so that the target and its surrounding information can be considered simultaneously.In the GM-APD Li DAR long-range vehicle data set,the accuracy of the proposed algorithm reaches 96.5%.Compared with other algorithms,the proposed algorithm has the highest accuracy.Finally,aiming at the problem that the detection based on intensity images is difficult in the complex scenes.This paper proposes an object detection algorithm based on information fusion.Firstly,the improved feature pyramid network which proposed in the previous section to enhance the accuracy of selection in the intensity images.Secondly,the intensity images and the range images are combined into point clouds with intensity information in the candidate regions.Finally,dynamic graph convolution network(DGCNN)is used to perform secondary detection of the target in the candidate regions.In the GM-APD Li DAR long-range vehicle data set,the AP of the network achieves 98.8%,and it has good robustness for complex scenes such as incomplete vehicle structure,weak echo and strongly reflected light spot,which provides a feasible solution for Li DAR dim object detection.