There are many small objects in traffic scenes, but due to their low resolution and limited information, their detection is still a challenge. Small object detection is very important for the understanding of traffic scene environments. To improve the detection accuracy of small objects in traffic scenes, we propose a small object detection method in traffic scenes based on attention feature fusion. First, a multi-scale channel attention block (MS-CAB) is designed, which uses local and global scales to aggregate the effective information of the feature maps. Based on this block, an attention feature fusion block (AFFB) is proposed, which can better integrate contextual information from different layers. Finally, the AFFB is used to replace the linear fusion module in the object detection network and obtain the final network structure. The experimental results show that, compared to the benchmark model YOLOv5s, this method has achieved a higher mean Average Precison (mAP) under the premise of ensuring real-time performance. It increases the mAP of all objects by 0.9 percentage points on the validation set of the traffic scene dataset BDD100K, and at the same time, increases the mAP of small objects by 3.5%.
18 April 2023 in IEEE Transactions on Intelligent Vehicles
This paper focuses on the problem of autonomous valet parking trajectory planning in complex environments. The task normally can be well described by an optimal control problem (OCP) for the rapid, accurate, and optimal trajectory generation. Appropriate initial guesses obtained by sampling or searching-based methods are important for the numerical optimization procedure. Still, in highly complex environments with narrow passages, it may incur high computational costs or fail to find the proper initial guess. To address this challenge, an enhanced hybrid A* (EHA) algorithm is proposed to address the issue. The EHA includes four steps. The first step is to quickly obtain the global coarse trajectory using a traditional A* search. The second step is constructing a series of driving corridors along the rough trajectory, then evaluating and extracting nodes from each wide or narrow passage based on the length of the box's side. The third step is to extract each passage's boundary points. The final step is connecting boundary points by hybrid A* and generating a feasible initial guess for OCP. To reduce safety risks, vehicles in particular areas (Fig. 1b) should travel as slowly as possible. The global and local speed restrictions are distinct, and local restrictions are only activated when the vehicle enters a particular area. This “if-else” structure makes the optimization problem difficult. A novel approximation formulation for these local state constraints is introduced to overcome this issue. The experimental results demonstrate that the proposed method for trajectory planning is effective and robust.
Trajectory prediction of heterogeneous road agents such as vehicles, cyclists, and pedestrians in dense traffic plays an essential role in self-driving. Despite breakthroughs in trajectory prediction technology in recent years, challenges remain in world state representation, social interaction modeling, real-time computing, and road agent heterogeneity. To address these challenges, we propose a new model that employs hierarchical convolutional networks and multi-task learning to predict agents' trajectories. The model first achieves effective and unified representation of agent and scene context by rendering heterogeneous world states in a top-down multi-channel raster map. Based on this representation, we propose hierarchical convolutional networks to extract global interaction and local features of all agents simultaneously, enabling the model to predict multiple agents' trajectories in real-time in a single forward inference. In addition, we specifically design multi-task learning branches with dynamic adaptive anchors to capture differences in behavioral patterns of heterogeneous agents, allowing a single model to accurately predict multimodal trajectories of multi-class agents. Extensive experiments on public nuScenes and Lyft datasets demonstrate top model performance. Importantly, our model is faster (2.2x) and more computationally stable than state-of-the-art models, making it well-suited for mass-produced self-driving systems that require both performance and computational efficiency.
31 July 2013 in Environmental Science and Pollution Research