基于双重注意力时空图卷积网络的行人轨迹预测

Pedestrian trajectory prediction based on dual-attention spatialtemporal graph convolutional network

当前行人轨迹预测研究面临两大挑战:1)如何有效提取行人前后帧之间的时空相关性;2)如何避免在轨迹采样过程中受到采样偏差的影响而导致性能下降.针对以上问题,提出基于双重注意力时空图卷积网络与目的抽样网络的行人轨迹预测模型.利用时间注意力捕获行人前后帧的关联性,利用空间注意力获取周围行人之间的相关性,通过时空图卷积进一步提取行人之间的时空相关性.引入可学习的抽样网络解决随机抽样导致的分布不均匀的问题.大量实验表明,在ETH和UCY数据集上,新方法的精度与当前最先进的方法相当,且模型参数量减少1.65×10^(4),推理时间缩短0.147 s;在SDD数据集上精度虽略有下降,但模型参数量减少了3.46×10^(4),展现出良好的性能平衡,能为行人轨迹预测提供新的有效途径.

There are two major challenges in current research on pedestrian trajectory prediction:1)how to effectively extract the spatial-temporal correlation between the front and back frames of pedestrians;2)how to avoid performance degradation due to the influence of sampling bias in the trajectory sampling process.In response to the above two problems,a pedestrian trajectory prediction model was proposed based on the dual-attention spatial-temporal graph convolutional network and the purposive sampling network.Temporal attention was utilized to capture the correlation between the front and back frames,and spatial attention was utilized to capture the correlation between the surrounding pedestrians.Subsequently,the spatial-temporal correlations between pedestrians were further extracted by spatial-temporal graph convolution.Meanwhile,a learnable sampling network was introduced to resolve the problem of uneven distribution caused by random sampling.Extensive experiments showed that the accuracy of this method was comparable to that of the current state-of-the-art methods on the ETH and UCY datasets,but the number of model parameters and the inference time were reduced by 1.65×10^(4) and 0.147 s,respectively;while the accuracy on the SDD dataset slightly decreased,but the amount of model parameters was reduced by 3.46×10^(4),which showing a good performance balance.The proposed model can provide a new effective way for pedestrian trajectory prediction.