*3.1.3 Analysis, relaxation of ideal tracking assumptions and conclusions*

In its current form, the work in [9, 10] uses the time series of the coordinates of the trajectories of airport travelers for deep learning. In the future, additional features could be exploited. Such features are the velocity, acceleration and heading of the traveler. Moreover, alternative deep learning architectures could be tested such as the ones that account for contextual anomalies [11]. Furthermore, experiments on realworld data of human trajectories should be conducted. Such data are expected to contain more subtle and sophisticated anomalies. Finally, procedures that degrade data quality and emulate more realistic operational conditions are being implemented in order to test our system in the artificial presence of missing data, noisy data, data association issues, as is the case with data capturing devices operating under realistic operational condtions. Nevertheless, the present work and framework allow security investment decisions on tracking devices and infrastructure to be made by assessing the effectiveness of such an investment through the proposed risk assessment method that envelops the performance of any such system from above by considering ideal tracking conditions through perfect knowledge of all agents' location. The proposed method and framework is currently being extended to cover other border security modalities, such as sea, land as well as multimodal crossing points in the context of the EU-funded TRESSPASS project [4].

In conclusion, a deep learning architecture for real-time risk assessment based on the trajectories of airport travelers as proposed in [9, 10] can be used for assessing risk without interrupting or delaying the flow of passengers at an airport or BCP at large. The architecture implements a deep RNN network and is fully automated. Thus, it is expected to be of great use to the human operators monitoring airport surveillance footages, reducing the potential errors and misjudges. The proposed risk assessment system is tested on a realistic, synthetic data set generated with the iCrowd simulator tailored to data sets representing traveler movements at the Luxembourg airport; however, any airport or BCP could have been modeled and used instead. The experimental results are very promising and they indicate that further security improvements at airport control points are achievable through risk assessment without inducing additional delays. This is due to the fact that the suspicious behavior threshold, derived by the deep learning procedure in [9, 10], lies at such a level so as to capture the malicious behavior while, at the same time, reducing false-positive alerts.
