Scene perception and trajectory forecasting are two fundamental challenges that are crucial to a safe and reliable autonomous driving (AD) system. However, most proposed methods aim at addressing one of the two challenges mentioned above with a single model. To tackle this dilemma, this paper proposes spatio‐temporal semantics and interaction graph aggregation for multi‐agent perception and trajectory forecasting (ST‐SIGMA), an efficient end‐to‐end method to jointly and accurately perceive the AD environment and forecast the trajectories of the surrounding traffic agents within a unified framework. ST‐SIGMA adopts a trident encoder–decoder architecture to learn scene semantics and agent interaction information on bird’s‐eye view (BEV) maps simultaneously. Specifically, an iterative aggregation network is first employed as the scene semantic encoder (SSE) to learn diverse scene information. To preserve dynamic interactions of traffic agents, ST‐SIGMA further exploits a spatio‐temporal graph network as the graph interaction encoder. Meanwhile, a simple yet efficient feature fusion method to fuse semantic and interaction features into a unified feature space as the input to a novel hierarchical aggregation decoder for downstream prediction tasks is designed. Extensive experiments on the nuScenes data set have demonstrated that the proposed ST‐SIGMA achieves significant improvements compared to the state‐of‐the‐art (SOTA) methods in terms of scene perception and trajectory forecasting, respectively. Therefore, the proposed approach outperforms SOTA in terms of model generalisation and robustness and is therefore more feasible for deployment in real‐world AD scenarios.

No abstract

Meshfree methods are suitable for solving problems on irregular domains, avoiding the use of a mesh. To deal with the boundary conditions, we can use Lagrange multipliers and obtain a sparse, symmetric and indefinite system of saddle-point type. Many methods have been developed to solve the indefinite system. Previously, we presented an algebraic method to construct an LU-based preconditioner for the saddle-point system obtained by meshfree methods, which combines the multilevel clustering method with the H-matrix arithmetic. The corresponding preconditioner has both H-matrix and sparse matrix subblocks. In this paper we refine the above method and propose a way to construct a pure H-matrix preconditioner. We compare the new method with the old method, JOR and smoothed algebraic multigrid methods. The numerical results show that the new preconditioner outperforms the preconditioners based on the other methods.In [2], a scheme based on smoothed algebraic multigrid (AMG) is proposed to solve the saddlepoint systems from meshfree methods. In this paper we present an H-matrix-based approach to solve the saddle-point system. Hierarchical matrices (H-matrices) were introduced in [3] and since then, much work has been done on the theory and applications of H-matrices [4][5][6][7][8]. Hmatrices provide a cheap but approximate way of carrying out matrix computations involving dense matrices.The basic idea of the H-matrix representation is that instead of representing a matrix exactly, approximations are used to represent a matrix based on a hierarchical block cluster tree. The leaves of the tree represent the matrix blocks that are not partitioned further, which are represented by low-rank matrices (Rk-matrix format) or by full matrices (full matrix format). An H-matrix arithmetic is developed by adapting conventional matrix operations to the H-matrix format. The required storage for a matrix and the computational complexity of operations, such as matrix-vector multiplication, matrix-matrix multiplication, matrix addition, inversion and LU factorization, are reduced to almost linear complexity (O(n log n)) [4,8].H-matrices have been combined with finite element methods to solve partial differential equations. Classic H-matrix construction [8] relies on the underlying geometric structure of the problem. Recently, a black-box H-matrix construction method was developed for sparse matrices, which is based on the matrix graph of the underlying sparse matrix [9].In [10], we present a pure algebraic method for the H-matrix construction, which is based on multilevel clustering methods. The experimental results show that it is comparable with the domaindecomposition-based H-matrix construction methods for sparse matrices. Since the algebraic method for the H-matrix construction does not depend on the geometric information underlying the problem, it can be applied to systems from meshfree methods. In [10], we present a way to build preconditioners for solving saddle-point systems, which have both H-matrix subblocks and sparse matri...

No abstract

No abstract