Intermodal Mobility Assistance for Megacities - Smart Information Systems: Computational Intelligence for Real-Life Applications - page 355

Table 13.1 Characteristics of the UCSD Highway Traffic Dataset

Level Description No. of videos

Light Free-flow traffic. Low number of vehicles at high speed 165

Medium Average number of vehicles at reduced speed 45

Heavy Stop-and-go traffic. High number of vehicles at very low speed 44

traffic surveillance camera over two days. In the dataset, each video is manually
annotated as light (i.e., free-flowing traffic), medium (i.e., traffic at reduced speed),
or heavy (i.e., stopped or very slow speed traffic) traffic. The dataset is challenging,
since a multitude of weather conditions are represented (e.g., clear, raining, and
overcast). Table 13.1 presents the main characteristics of the dataset in more detail.

13.3.2.5 Experimental Setup

OpenCV Java 7 is used to implement the extraction of optical flow vectors and to
generate the traffic flow analysis SVM model. Motion descriptors are constructed
based on optical flow vectors as explained in Sect. 13.3.2.2 . We trained the multiclass
SVM model with an Radial Basis Function (RBF) kernel. SVM parameters were
optimized by fivefold cross-validation on the training data.
We adopted the same training and testing methodology as [ 5 , 24 ]. We repeated
the tests four times with different training and test samples, where in each repetition
the dataset was split with 75 % for training and cross-validation and 25 % for testing.

13.3.2.6 Results and Discussion

In this section, we present the classification performance of our method together
with the related confusion matrices. We also compared our SVM-based method with
other classification schemes such as k-NN, Naive Bayes, and AdaBoost learning
methods. The classification performance of each method is shown in Table 13.2 .
We achieved 94.90 % classification accuracy on average with motion vector-based
representations and multiclass SVM on the UCSD dataset. In addition, the multiclass
SVM outperformed other classification methods such as k-NN, Naive Bayes, and
AdaBoost.
In Fig. 13.5 , the confusion matrix of the classification results of our method for
the UCSD dataset is illustrated. The confusion matrix represents the performance of
our method with motion vector-based representations using a multiclass SVM. The
detailed definition of the labels presented in Fig. 13.5 is given in Sect. 13.3.2.4 .As
illustrated in Fig. 13.5 , medium - heavy pairs are the most confused congestion-level
label pairs. This result suggests that there is a need to incorporate additional visual

7

http://opencv.org/opencv-java-api.html .