What did you find out?

In our research, we examined the performance of various algorithms for short-term pedestrian trajectory prediction, focusing on their robustness, accuracy, and runtime efficiency, particularly for use in autonomous vehicle systems. We found that simple models, like the Constant Velocity Model (CVM), remain competitive with more complex AI-based approaches and even outperform them in certain scenarios. Our analysis revealed that for predicting pedestrian behavior, the last second is crucial. This is because, especially in complex environments like traffic, people quickly change their behavior and continuously adapt to dynamic situations.

What challenges did you face during the research? 

Although many algorithms are available as open-source, simplifying the analysis of existing methods, preparing a given model for use can vary in complexity. Application-specific aspects are often not sufficiently explained in related publications, leaving out details on robustness, computational efficiency, or accuracy. Additionally, the evaluation methodology used often doesn’t target real-world application, resulting in the selection of trajectories with the least error rather than the most probable one. We also had to rely on the popular ETH/UCY pedestrian dataset, which includes only pedestrians on sidewalks. Therefore, the models we studied are difficult to generalize to urban scenarios, as they lack critical semantic information like road infrastructure and social interactions with vehicles.

What are the practical applications of your research?

Our findings are essential for developing the next generation of application-oriented prediction algorithms, such as those used in driver assistance systems or autonomous vehicles. Our experiments will help in the future to selectively choose relevant information for AI-based methods, making them safer and more efficient. Furthermore, our results indicate that simpler (non-AI) models can be used in many scenarios, which require fewer computational resources in mobile systems. Finally, we believe that categorizing pedestrian actions into "discrete actions" like walking or stopping might be more effective than focusing solely on their movement. Thus, the field of prediction still offers plenty of opportunities and space for creative ideas.

Publication

Evaluating Pedestrian Trajectory Prediction Methods With Respect to Autonomous Driving; Nico Uhlmann, Felix Fent, Markus Lienkamp; IEEE Transactions on Intelligent Transportation Systems; 2024