Study On Pedestrian Trajectory Prediction For Intelligent Vehicles

The adoption of information fusion and deep learning techniques has greatly boosted the development of intelligent vehicles.With remarkable performance of environment perception,current intelligent vehicles are capable of recognizing and localizing objects around the vehicles in acceptable accuracy and velocity.To navigate in sophisticated urban roads,however,intelligent vehicles still face challenges in understanding the intentions of other road users.To tackle this problem,an appearance-based deep learning method for pedestrian trajectory prediction is proposed in this work.The main research contents of this work are summarized as follows:A deep learning model based on encoder-decoder architecture is built that utilizes pedestrian appearance as well as historical locations and camera ego-motion to predict future trajectories of pedestrians for vehicle applications.The proposed model consists of three encoders and a decoder,namely: appearance encoder,location encoder,and ego-motion encoder,to deal with different types of inputs;and a Long Short-Term Memory(LSTM)decoder.A disparity-guided attention mechanism is designed in the appearance encoder,which attends to salient dynamic regions in pedestrian appearance and reduces redundancy.To avoid overfitting problem of extra trainable parameters,the proposed attention mechanism does not contain any trainable parameters.The proposed model is implemented in Tensor Flow.Two public datasets are chosen to conduct training,tuning,evaluation and visualization.Experiment results show that the proposed model outperforms the baselines in predicting accuracy and has interpretability.For further validation of the generalization ability of the proposed method,real-world pedestrian data are collected and automatically labelled with open-source object tracking framework Deep SORT.The proposed model is evaluated and compared with the baselines on the collected data.