Conv-LSTM: Pedestrian Trajectory Prediction in Crowded Scenarios

Pedestrian trajectory prediction is a challenging problem in the crowded and chaotic scenarios. Currently, the prediction error is still high because the input of Long Short-Term Memory (LSTM) network is a 1D vector, which cannot represent the spatial information of pedestrians. To tackle this, we propose to use tensors to represent the complex environmental information. Meanwhile, LSTM internal full connection is converted into full convolution to predict the spatiotemporal pedestrian trajectory sequences. The results show that our method reduces the displacement offset error better than recent works including Social-LSTM, SS-LSTM, CNN, Social-GAN, Scene-LSTM, providing more realistic trajectory prediction for the chaotic crowd.