Multi-Object Tracking Based On Interactive Graph Neural Network

Graph is the most convenient structure to represent the association between a group of target nodes.Meanwhile,graph neural network has been studied deeply in recent years,and some researchers have applied graph to the automatic driving of vehicles in recent years.However,most researchers carry out simple graph construction,such as K-NN graph,fully connected graph,etc.,which either depends on the selection of parameters or specific data sets,or causes high information redundancy,making the model difficult to train.In the autonomous driving of vehicles,the safety of the system is placed in the first place,and the reliability and interpretability of the system are also very important.As the basic task of automatic driving,multi-target detection and tracking is the focus of this paper.In order to make reasonable use of node information and association,this paper proposes an interactive graph neural network of multi-target tracking,which can accurately identify and track multiple categories and multiple targets in complex driving environment.In this paper,Multi-Object Tracking based on Interactive Graph Neural Network is proposed for multi-object tracking tasks in complex scenes.Firstly,the point cloud is regularly divided into pillars of the same size in space,and the point cloud pillars are sampled at different densities.Feature extraction is carried out on the sampling results to obtain the feature pseudo-graph.Secondly,the Attention Region Proposal Network extracted high confidence Proposals from the feature pseudo graph.The Proposals were taken as graph nodes and the relationship between object feature similarity and location was constructed.Then feature learning and updating of graph nodes are carried out by multi-layer perceptron.Finally,multi-detector is used for detection and tracking.Experimental results show that the Multi-Object Tracking based on Interactive Graph Neural Network proposed in this paper has better learning ability for different categories,and the Average Multiple Object Tracking Accuracy is 16.8% and 1.7% higher than that of the latest method of the same type in Nu Scenes data set.In order to solve the unsatisfactory tracking effect in object tracking for incomplete global information acquisition of Multi-Object Tracking based on Interactive Graph Neural Network in target tracking task,an improved Graph Neural Network based on self Attention and Interactive of Multi-Object Trackiing is proposed in this paper.In the updating of graph nodes,not only the feature of nodes but also the proportion of each node influencing each other are considered.The influence weight of neighbor node and object node is determined by the correlation between object node and neighbor nodes,and the importance of nodes is considered from various perspectives.The state of a node is updated by a weighted sum of the state of the node itself and all of its neighbors.Experimental results show that the improved algorithm can obtain more comprehensive node information easily,and the detection result and tracking result are 22.15% and 10.78% higher than those of the same algorithm.