04/2019 – 10/2019
A good description of the environment is a benefit for positioning, navigation systems and autonomous driving. One of the most original but important environment description tools is a map, which describes the physical environment using topological models and known metrics. Generally, maps can include both geometric (the building’s layout and the locations of blocking objects) and semantic (the locations where specific actions tend to take place, e.g. bike road, road driving directions, …) components.
Both of these components can be used to enhance positioning methods by providing geometric limitations and intention-based prediction models. For example, in an open street, there are multiple different agents (e.g. pedestrians, cyclists, cars etc.), which tend to move in a specific and distinguishable manner, based on the traffic rules, geometric constraints, and characteristic movement patterns. Hence, the classification and prediction of the agents’ movements in specific environments can be used to enhance positioning, e.g. in 5G positioning networks, to control and optimize traffic, and to aid with autonomous driving scenarios. Learning a geographical representation of the agent behavior can provide situation-based motion models for agent groups, point out security risks, and provide analysis of the behavioral trends in specific environments. With the increased availability of positioning solutions and network deployment, especially with the arrival of the 5G standard, higher amounts of trajectory data will be available for different environments, making trajectory-based data analysis as an application more feasible .
While these kinds of problems have been studied for years, recently some interesting both supervised and unsupervised approaches have been proposed . Zheng et al  proposed a method to infer pedestrian motion models based on point-based segmentation and a decision-tree based inference model. Social generative adversarial nets (GANs)  have also been proposed for trajectory data analysis with the aim to predict socially plausible futures of human motion paths by adversarial training against a recurrent discriminator, and encourage diverse predictions with a novel variety loss. Unsupervised clustering methods include the construction of geometric maps from trajectory data by clustering sets of relative sub-trajectories  or line segments .
In this thesis, the overall goal is to use trajectory data to learn environment maps that describe the movement patterns of agent groups in specific environments. The maps will include both geometric constraints and agent groups with different movement patterns and intentions. For the development and evaluation of the developed methods, the Stanford Drone dataset  is used, a large dataset containing annotated videos of different crowded outdoor scenarios with different kinds of agents (pedestrians, cyclists, skateboarders, cars, busses …). First, the agent groups are to be classified based on the characteristic features of the trajectories. The classification algorithms are to be evaluated using established metrics like confusion matrices, accuracy and recall. After that, the characteristic behavior of agent groups in the different environments is modeled as a spatial distribution or map representing motion peculiarities of the groups at different spots, like the likelihood of using specific routes (like a bike lane), the characteristic velocities, accelerations and angular rates at different spots (e.g. pedestrians slowing down before they cross a street). The feasibility of the representation is to be evaluated by checking how well the modeled distribution represents a separate test dataset using statistical metrics like the Kullback-Leibler divergence, or similar methods depending on the chosen algorithm for learning the spatial representation.
- Wan Y., Zhou C., Tan H., Pei T.: Semantic-Geographic Trajectory Pattern Mining Based on a New Similarity Measurement ISPRS Int. J. Geo-Information 6 (2017): 212.
- Bian J., Tian D., Tang Y., Tao D.: A survey on trajectory clustering analysis. arXiv:1802.06971v1 [cs.CV] 20 Feb 2018
- Zheng Y., Xie X., Ma W.: Understanding Mobility Based on GPS Data. Proceedings of the 10th ACM conference on Ubiquitous Computing (Ubicomp 2008)
- Buchin K. ; Buchin M., Duran D., Fasy B., Jacobs R.; Sacristan V., Silverira R., Staals F., Wenk C.: Clustering Trajectories for Map Construction. 25th International Conference on Advances in Geographic Information Systems, 2017
- Agrim G., Justin J., Li F., Silvio S., Alexandre A.: Social GAN: Socially Acceptable Trajectories with Generative Adversarial Networks.org/abs/1803.10892
- Lee J., Han J., Whang: Trajectory Clustering: A Partition-and-Group Framework. DOI: 10.1145/1247480.1247546
- Robicquet A., Sadeghian A., Alahi A., Savarese S.: Learning Social Etiquette: Human Trajectory Prediction In Crowded Scenes. European Conference on Computer Vision (ECCV), 2016