【摘要】 行人轨迹预测在自动驾驶和监控系统等多个应用中具有重要意义.目前大多数行人轨迹预测模型采用基于循环神经网络的编码器-解码器结构，其自回归的解码结构存在一定的累积误差，而且循环神经网络对序列的长期依赖问题仍然无法很好地解决.本文提出一种基于Transformer网络的非自回归行人轨迹预测模型，非自回归的解码结构能够同时生成所有预测值来减少累积误差，Transformer网络中的自注意力机制能够改善长期依赖问题.本文还设计一个局部信息加强模块来捕获行人运动趋势发生变化的局部特征，同时结合边界框的位置信息和大小信息来编码第一视角下透视投影产生的影响，使得模型提取到的轨迹特征更加有效.实验结果表明，在基于第一视角的公开数据集PIE(Pedestrian Intention Estimation)上，本文提出的模型比PIE预测模型在15、30、45帧的平均位移误差和终点位移误差上分别降低了24%,14.5%,11%和6%.更多还原

【Abstract】 Pedestrian trajectory prediction plays an important role in many applications such as automatic driving and monitoring systems. At present, most pedestrian trajectory prediction models are recurrent neural network(RNN) based on encoder-decoder architectures. RNN could not solve the long-term dependence, and its auto-regressive decoding scheme introduces accumulate errors. This paper proposes a Transformer based non-autoregressive pedestrian trajectory prediction model, whose non-autoregressive decoder can generate all predictions simultaneously to reduce accumulative errors. The self-attention mechanism can enhance the long-term dependence problem. More specifically, this paper designs a local information enhancement module to extract the local features when pedestrian’s movement trend changes, and combining with the location information and scale of the boundary encodes the impact of perspective projection in the first perspective,which makes the trajectory features extracted from the model more efficient. Experimental results show that, compared with the PIE(Pedestrian Intention Estimation) model, the average displacement error of 15, 30 and 45 frame and the end displacement error are respectively reduced by 24%,14.5%,11% and 6% on a public data set PIE based on the first perspective.更多还原