Lane change and crash risk prediction are critical technologies for autonomous driving. An attention-based LSTM model is proposed in this paper for lane change behavior prediction in highways, considering both the surrounding vehicles’ information around the target vehicle at the current time and the historical trajectory data of the vehicle, and it shows higher accuracy and better interpretability than other models. There are two sections in this prediction model, including a pre-judgment model based on the C4.5 decision tree and bagging ensemble learning, and a multi-step lane change prediction model of LSTM with attention mechanism. In addition, the model is applied in the NGSIM datasets to test practicability and accuracy, and the precision of left lane change and right lane change is 98% and 94% respectively 1 second before the vehicle changes lane. Finally, to judge the safety of vehicles in the driving process, a crash risk prediction model based on Time-To-Collision is proposed, which further verifies the effectiveness of the method.

The game of GO has long been regarded as the most challenging game for artificial intelligence because of its enormous search space and the difficulty of evaluating its board positions. In early 2016, the defeat of Lee Sedol by AlphaGo became the milestone of artificial intelligence. AlphaGo’s success lies in that it efficiently combines policy and value networks with Monte Carlo tree search (MCTS). And these deep convolutional neural networks (DCNNs) are trained by the combination of supervised learning and reinforcement learning. However, large convolution operations are computationally-intensive and typically require a powerful computing platform, for example, a graphics processing unit (GPU). Therefore, it is challenging to apply DCCNs in resource-limited embedded systems. Field programmable gate array (FPGA) is proposed to be an appropriate solution to implement real-time DCCNs models. However, the limited bandwidth and on-chip memory storage are the bottlenecks for DCCNs acceleration. In this paper, an AlphaGo Policy Network is designed, and efficient hardware architectures are proposed to accelerate the DCCN model. The accelerator can be fit into different FPGAs, providing the balancing between processing speed and hardware resources. As an example, the AlphaGo Policy Network is implemented on Xilinx design suite VCU118, and the results show that our implementation achieved a performance of 3036.32 GOPS and achieved up to 56x speedup compared to CPU and 22.4x speedup compared to GPU.

Accurately forecasting the motion of surrounding vehicles is a crucial prerequisite for achieving safe autonomous driving. Methods for trajectory prediction encompass both physics- and learning-based approaches. Despite the significant advancements made by learning-based methods in enhancing performance, ensuring that predicted trajectories are realistic, interpretable, and physically feasible remains a challenging problem. In this study, we propose a physics-based deep learning framework founded on an encoder-decoder architecture, modeling the historical motion of traffic agents effectively coupled with the surrounding environment information through attention mechanisms. Our model incorporates the vehicle dynamic model to couple with vehicle motion and utilizes the Kalman filter to fuse scene context information for multi-step accurate and feasible multimodal trajectory prediction. Our method leverages the strengths of both learning- and physics-based models. The extensive experiment results on the Lyft l5 dataset demonstrate that our model outperforms various baseline approaches in terms of metrics including prediction errors, feasibility, and explainability.

Despite the ability of 3D convolutional methods to extract spatio-temporal information simultaneously, they also increase parameter redundancy and computational and storage costs. Previous work that has utilized the 2D convolution method has approached the problem in one of two ways: either using the entire body sequence as input to extract global features or dividing the body sequence into several parts to extract local features. However, global information tends to overlook detailed information specific to each body part, while local information fails to capture relationships between local regions. Therefore, this study proposes a new framework for constructing spatio-temporal representations, which involves extracting and fusing features in a novel manner. To achieve this, we introduce the multi-feature extraction-fusion (MFEF) module, which includes two branches: each branch extracts global features or local features individually, after which they are fused using multiple strategies. Additionally, as gait is a periodic action and different body parts contribute unequally to recognition during each cycle, we propose the periodic temporal feature modeling (PTFM) module, which extracts temporal features from adjacent frame parts during the complete gait cycle, based on the fused features. Furthermore, to capture fine-grained information specific to each body part, our framework utilizes multiple parallel PTFMs to correspond with each body part. We conducted a comprehensive experimental study on the widely used public dataset CASIA-B. Results indicate that the proposed approach achieved an average rank-1 accuracy of 97.2% in normal walking conditions, 92.3% while carrying a bag during walking, and 80.5% while wearing a jacket during walking.

Ore segmentation and proportion detection (PD) in conveyor belt images (CBI) are significant for the safe operation of ore engineering. However, the actual accuracy of segmentation and detection is low due to the influence of ore flow shape, irregular distribution, and dust interference. This paper proposed RSFA-Net, a high-performance and lightweight network for segmentation and PD in ore conveyor belts (OCB). Firstly, a new double-layer residual depth-separable (DRDS) algorithm is proposed for the first time and applied to encoder-decoder networks. The module improves the ability of the proposed network to identify ores in images and represent complex features, while greatly reducing the parameters and computational burden required to segment the network. Secondly, we also propose a new channel spatial fusion attention (NCSFA) to fully tap the potential of the proposed network for ore feature extraction, which can capture complex and variable ore features and extract more critical information in the context of the image. Finally, we used high-performance image sensors to build a large, high-quality OCB dataset, solving the problem of the unavailability of public datasets. We proposed and adopted the maximum relative error (MRE) as the precision index of the ore proportion detection task for the first time. RSFA-Net was compared to nine of the most representative ore segmentation networks. RSFA-Net outperforms all methods in terms of segmentation performance and lightweight. The segmentation accuracy of RSFA-Net is 97%, while the model size is only 1.06MB, and the accuracy of PD is as high as 98%.