**Figure 17.**

*Confusion matrix.*


#### **Table 1.**

*3.1.2 Analysis and relaxation of ideal tracking assumptions - experimentation*

tions make the aforementioned phenomenon even more intense.

tive effect on the detection of abnormal trajectories as discussed next.

Most of the *false negatives* are abnormal trajectories that cannot easily be discriminated from the normal even by a human operator. Soft thresholding could be used in order to raise alerts for a human supervisor. On the other hand, the main reason for the *false positives* is the fact that airport travelers chose to move in ways that may not necessarily be similar to the normal trajectories. Large airport conges-

Although the conditions the risk assessment algorithm was evaluated under assumed perfect knowledge of the traveler trajectories, relaxation of the assumption of perfect knowledge of the traveler trajectories by injecting noise in the position accuracy and/or assuming missing position data, did not have a considerable nega-

In order to assess the performance of the anomaly detection algorithm in realistic conditions we introduce noise in the data to emulate the uncertainty in passengers' positions reports. The "noisy data" emulate the inaccuracy in the reports of the

*with noisy data*

**Figure 16.** *ROC curve.*

**128**

**Figure 15.**

*Precision-recall diagram.*

*Deep Learning Applications*

*Values of the recruited evaluation measures.*

positions of the people in the space. Under realistic conditions, the tracking and risk assessment system will receive data from inaccurate sources, such as cameras, sensors, etc. used to estimate distances, mobile signal strength, etc.

In stark contrast, the iCrowd emulator produces people and their movements, and periodically reports the exact ones (so without noise) their positions in the risk assessment system. During the preprocessing of this data the possibility of the system to add Gaussian noise, the "volume" of which (parameter σ2 of Gaussian noise) is given by the user. This is obviously not intended to never be used in real application, and exists only for experimentation. For examining the behavior of the system under realistic conditions is required noisily data of different intensity.

Noise can enter the system in 2 cases: during training and during testing or actual application. It is known that when training any neural network, it is good to have variety in the data in which the network is exposed so that it is not overtrained. So, it is expected that training with Noisy data can improve the overall performance of the system. During testing or the actual implementation of the system would definitely be better to have perfect data, but unfortunately this is often impossible. In the context of the internship training and validation data were performed with Gaussian noise with σ2 from 0 to 1.9 with step 0.1, testing data with corresponding noise levels, and for each combination they were trained and evaluation of the neural network, and metrics were calculated for each of them. The metrics used were the Receiver Operating Characteristic curve (ROC curve), the Precision-Recall curve (PR curve), and the corresponding Area Under Curve (AUC) scores. These metrics give similar results, in the sense that they are defined

in [0,1] with value range [0,1], so the minimum AUC value is 0 and the maximum is 1. Higher value means better true positive and true negative to false positive and false negative ratio. The two metrics generally return similar results, but in our case more weight is given in metric PR, as it offers a better estimate in cases that interest us more the positive class of results, or the results consist of significantly more elements of one of the two classes. Both features apply in the case of this risk assessment system.

During the experiment, networks emerged that failed to find an acceptable solution to why either they were trapped in a local minimum or they encountered the phenomenon of exploding gradient. These cases appeared to be random and independent of the parameters noise, so the network was initialized differently and the training started from the beginning. The experiments were performed using 3 levels of congestion in space, low, moderate, and high, and for each of them 40 neural networks were created, one for each training/testing noise combination mentioned above. For every desired network, 4 independent trainings were conducted and the averages of metrics of interest were kept. The same test data, corresponding to low-to-medium congestion, were used for testing all tested models. The final results are presented below (**Figures 18**–**20**):

From the ROC AUC score graphs above, it is seen that the models that result in the highest performance correspond to the following noise level in the training data:

For low noise data, the best performing data with AUC = 0.91 corresponds to noise level σ<sup>2</sup> = 0.8 in the training data.

For medium noise data, the best performing data with AUC = 0.96 corresponds to noise level σ<sup>2</sup> = 1.6 in the training data.

For high noise data, the best performing data with AUC = 0.95 corresponds to noise level σ<sup>2</sup> = 1.4 in the training data.

From the above results it is clear that the performance of the networks remains constant when we apply noise to the test data. This implies that, since training completed, the network remains robust and is not affected by data noise, so it can to be used in a real application. Of particular interest are variations that occur when present noise in education data. As mentioned above, training a neuron network usually benefits from the difference in training data, as it helps learn the patterns that appear in the data instead of the data itself. This obviously does not mean that the more noise the better. In every network and for every application there is some optimal noise level that offers the best performance. At cases with low and

> moderate congestion it seems that the Gaussian noise with σ2 0.5–0.8 has the best performance, while for high congestion the training with noise performs better with σ2 1.4. The variance has not yet been attributed to any of its specific features

*Training with high congestions data with different noise levels. Testing with low-to-medium congestion data.*

*Training with medium congestions data with different noise levels. Testing with low-to-medium congestion*

*Risk Assessment and Automated Anomaly Detection Using a Deep Learning Architecture*

*DOI: http://dx.doi.org/10.5772/intechopen.96209*

The evaluation results from the performance of the risk assessment algorithm with the iCrowd simulator demonstrates that risk assessment can be done accurately and without necessarily inducing additional delays in the security screening process since the trajectory classification in normal or suspicious is done by overhead cameras while the travelers go about their normal check-in routine at the airport. To that extent, the proposed risk assessment method based on anomaly detection on traveler trajectories can be used to improve the security screening effectiveness while keeping the delay low (or moving the operating point in

Furthermore, the proposed method can be used as a financial investment tool for estimating the cost of acquiring the necessary equipment (in this case overhead cameras) for a certain level (probability of accuracy) before purchasing it, and for performing a trade-off analysis between the cost of acquisition of the necessary

network, model, training method, or data.

**Figure 19.**

**Figure 20.**

**131**

*data.*

**Figure 10** from high delay to low.).

*Training with low congestions data with different noise levels. Testing with low-to-medium congestion data.*
