基于多特征融合的行人意图以及行人轨迹预测方法研究

行人作为重要的交通参与者，其行走意图和轨迹预测对智能驾驶汽车的决策规划具有重要意义。基于注意力机制增强的长短时记忆(Long Short-term Memory，LSTM)网络，设计一种多特征融合的行人意图以及行人轨迹预测方法。该方法通过融合行人骨架和头部方向特征，以加强行人运动特征的表达，并将融合特征作为意图预测网络输入，继而得到行人意图；由于行人运动具有不确定性，将行人意图预测类别和历史轨迹坐标的联合向量作为行人轨迹预测网络的输入，以期生成更为精确的轨迹预测结果。此外，在轨迹预测网络中引入注意力机制，以加强LSTM对各个时刻编码向量的有效利用，从而提高网络的行人轨迹预测性能，并基于Daimler数据集进行训练及验证。研究结果表明：所提出的多特征意图预测方法准确率可达96.0%，优于基于骨架单特征的意图预测网络；在预测时域为1 s的情况下，预测轨迹的位置均方根误差为347 mm，相较于恒速度(Constant Velocity，CV)模型、交互多模型(Interacting Multiple Model，IMM)、常规LSTM等基线方法均有明显的提升；在实际场景分析中，提出的方法可提前0.56 s识别行人的转弯意图，可为智能车辆的决策模块提供有益线索；提出的方法能够有效降低行人意图转变过程中的轨迹预测误差，对减小车辆与行人碰撞事故，提高智能车辆行驶安全性具有重要意义。

Prediction of Pedestrian Intention and Trajectory Based on Multi-feature Fusion

Pedestrians play an essential role as transportation participants. Predicting their intentions and future locations is crucial for the decision-making and planning of intelligent vehicles. In this study, a long short-term memory (LSTM) network with an attention enhancement mechanism was used to design a method based on multi-feature fusion for predicting the intention and trajectory of pedestrians. This method improves the representation of pedestrian motion by fusing the features of pedestrians' skeletons and head directions, and the fused features are used as the input of the intention-prediction network to obtain the pedestrian intention. To reduce the impact of pedestrian-movement uncertainty on trajectory-prediction performance, the joint vector of the pedestrian intention and the historical trajectory coordinates were used as the input of the trajectory-prediction network, which helped obtain more accurate prediction results. In addition, an attention mechanism was introduced into the trajectory-prediction network to strengthen the effective use of the LSTM coding vectors at each moment, which further improved the pedestrian-trajectory prediction performance. The Daimler dataset was used to train the proposed model. The test results show that the proposed multi-feature method has an accuracy rate of up to 96.0% for intention prediction, which is superior to that of the single-feature network, based on the pedestrian skeleton. The root mean square error of the position in the trajectory prediction is 347 mm with a prediction horizon of 1 s, which is better than that of baseline methods such as the constant-velocity model, interacting multiple models, and conventional LSTM. In an actual scene analysis, the proposed method can recognize the turning intention of pedestrians 0.56 s in advance and can provide useful clues for the decision-making subsystem of an intelligent vehicle. Finally, the proposed method can effectively reduce the trajectory-prediction error when dealing with changes in pedestrian intention. This will help reduce the occurrence of collisions between the ego vehicle and pedestrians, eventually improving the driving safety of intelligent vehicles.