Multi-information-based convolutional neural network with attention mechanism for pedestrian trajectory prediction
【Abstract】
We propose to combine the pose and the 2D-3D size information of the pedestrian to indicate human intention.
We propose to use the depth map to extract human-human interactions and human-scene interactions for future trajectory prediction.
We design a novel convolutional network model combining human intention, historical trajectories, human-human interactions, and human-scene interactions to predict the pedestrian's future trajectory. Meanwhile, compared with traditional methods based on LSTM, our model can speed up the calculation process through the parallel computing feature of convolution.
Predicting pedestrian trajectory is useful in many applications, such as autonomous driving and unmanned vehicles. However, it is a challenging task because of the complexity of the interactions among pedestrians and the environment. Most existing works employ long short-term memory networks to learn pedestrian behaviors, but their prediction accuracy is not good, and their computing speed is relatively slow. To tackle this problem, we propose a multi-information-based convolutional neural network (MI-CNN) to incorporate the historical trajectory, depth map, pose, and 2D-3D size information to predict the future trajectory of the pedestrian subject. After training, we evaluate our model on crowded videos in the public datasets MOT16 and MOT20. Experiments demonstrate that the proposed method outperforms state-of-the-art approaches both in prediction accuracy and computing speed.
【Author】
RuipingWanga, YongCuia, XiaoSongb, KaiChena, HongFangc
【Keywords】
Depth map, Pose, 2D-3D size information, Convolutional neural network, Trajectory prediction 
【Journal】
Image and Vision Computing(IF：2.7) Time：2021-01-24
