Assessing the lack of context knowledge for a pedestrian predicting neural network

Ensuring a safe journey with an autonomous vehicle, the surrounding has to be sensed and understood. Especially human intuition about the plans and intentions of traffic participants is hard to model for machines. In literature, there are already several prediction techniques existing for pedestrians, which are based on different features. Some models are very complex, whereas others only rely on the considered person’s motion. The goal of this work is to analyze the importance of different classes of context knowledge for the prediction performance, derive features to remove this lack of information and prove this by an improved prediction algorithm. In order to judge the lack of context knowledge, we analyze the prediction performance and error cases of a long short-term memory (LSTM) Neural Network as State-of-the-Art prediction algorithm, only based on motion data. The Network is trained and evaluated on a benchmark dataset, to make the results comparable to other approaches. Analyzing the most error-prone predictions, the missing context shall be identified, which could improve the prediction results. Since the data was generated by video, we can evaluate the whole scenario and identify the influencing factors. The found influences were classified in categories and their importance for the prediction model estimated. We prove the necessity of additional context knowledge by retraining a neural network with additional context knowledge. In a literature research we compare our found results to existing approaches.