3D Object Tracking And Detection Based On Feature Fusion And Siamese Network

The 3D point cloud is not affected by light,shadow,and other factors,and is an important data source for autonomous vehicles to perceive the vehicle environment.However,point cloud data’s sparseness makes accurate object tracking of 3D point clouds a challenging task.This work focuses on the direct use of point cloud information and proposes a more practical3D object tracking framework.Given the position,direction,and size information of the object in the first frame,the object’s 3D tracking task can be completed for the subsequent continuous point cloud frames.The following are introduced from three aspects.This paper proposes a Re-Track Framework for 3D point cloud tracking,which can re-track the lost object.Using the classic motion estimation algorithm,it completes the coarse stage with the proposal network.Ensuring accurate score feedback required by the Kalman filter helps to estimate the mean and variance of the Gaussian distribution of the potential bounding box and achieve good motion estimation.Therefore,the mean and variance of the Gaussian distribution obtained in the coarse stage,which are more accurate than that before the coarse stage,are shared with the fine stage.The loss of the tracking object triggers the fine stage.It uses the Kalman filter to receive the score feedback from the coarse stage.And the tracking result in the previous T frames helps to obtain the score feedback and improves the accuracy of the predicted Gaussian distribution’s mean and variance.Finally,the output box is selected with the highest cosine similarity among the 147 candidate boxes generated from the previous frame.Also,expanding the search range of the fine search stage helps obtain accurate and reliable tracking results.In terms of network structure,this paper proposes a Siamese Dense Autoencoder,it is suitable for object tracking and shape completion.This network structure adds a dense concatenation to the encoder-decoder structure,therefore realizing feature reuse.Simultaneously,inspired by SENet,the SE block is introduced at each layer of the encoder to filter the concatenated features from the channel.These designs will not greatly increase the number of network parameters.In terms of completeness,which is an index for evaluating network completion performance,the completion performance of the Siamese Dense Autoencoder is better than that of the public benchmark.The tracking performance is also better.As for feature fusion of temporal point clouds,this paper proposes a sample update strategy for point clouds to improve the model shape’s representation,which solves the tracking model update problem in the Siamese Network Tracker.This strategy calculates the chamfer distance between two samples to measure the difference between them.Specifically,the model update strategy generates different sample groups,the samples of the sample groups are similar,and the differences between the sample groups are large.Each time there is a new sample,a new sample group will be initialized.If the number of sample groups exceeds the limit L_N,the group with the smallest weight is discarded.Otherwise,we combine the samples with the smallest difference.According to the difference of samples,this strategy effectively combines small difference samples into the same sample group.Generally,the samples in consecutive frames are very similar.The model update strategy can keep the sample diversity while ensuring that the tracking process is not affected by the number of similar samples.