**5. Experimental results**

#### **5.1 Evaluation of the obstacle extraction module**

The first proposed module is evaluated on real-world data in outdoor environments in various weather conditions. The datasets concern videosurveillance system in which a stationary camera monitors real-worlds, such as a level crossing. The four datasets include: "Pontet" and "Chamberonne", which are two level crossings in Switzerland in cloudy weather, given that test images were 384 × 288 pixels; a dataset entitled "Pan", which represents a level crossing in France in sunny weather, given that test images were 720 × 576; and a dataset taken in snowy weather in EPFL–Switzerland is also considered for the evaluation. The test images are the same size as the "Pontet" and "Chamberonne" datasets. For a qualitative evaluation purpose, 1000 of foreground ground truths images have been obtained by manual segmentation from the "Pontet" and "Chamberonne" datasets. This allows computing the recall and the precision of the detection. In the experiments, the proposed framework is compared with the Mixture Of Gaussians (MOG) and Codebook algorithms. Furthermore, ICA model is evaluated on

obtained by Equation 23 and 24:

on Stereo Vision for Level Crossings Safety Applications

given by *Recall* and *Precision* measures:

Table 2. Processing time of Pontet dataset.

**5.2 Evaluation of the 3D localization module**

given by fugure 6.

*Recall* <sup>=</sup> *Tp*

<sup>93</sup> Intelligent Surveillance System Based

*Precision* <sup>=</sup> *Tp*

where *Tp* represents the number of well pixels classified as foreground, compared to the ground truth, *Fn* is the number of pixels classified as background, whereas they are really foreground pixels while referring to the ground truth, and *Fp* is the number of pixels classified as foreground, whereas they are really background pixels. *Tp* + *Fn* can be seen as the number of the true foreground pixels obtained by the ground truth, while *Tp* + *Fp* is the foreground pixels classified by a given algorithm. The image samples used for computing these two previous parameters are taken from the two datasets *Pontet* and *Chamberonne*, given that five hundred images from the each dataset are used for a manual extraction of foreground objects. This allows obtaining a ground truth dataset from which the different algorithms are evaluated. Table 1 shows the qualitative evaluation of the foreground extraction process,

*Tp* + *Fn*

*Tp* + *Fp*

**MOG Codebook ACI+Filtering**

*Recall* 94.76% 93.49% 96.14% *Precision* 95.87% 91.72% 97.34%

A visual comparison of our method compared with two other methods from the literature is

The implementation of the proposed framework runs on a personal computer with an Intel 32-bit 3.1-GHz processor. For the Pontet dataset, the proposed algorithm runs at a speed of 13 fps (frame per second). The processing time of our algorithm is compared with MOG and Codebook algorithms. Table 2 shows that our algorithm is faster than the other algorithms.

> **Algorithms** Proposed algorithm MOG Codebook **Processing time** 88.687 *ms* 286.588 *ms* 118.402 *ms*

The proposed depth estimation for 3D localization algorithm is first evaluated on the Middlebury stereo benchmark (http://www.middlebury.edu/stereo), using the Tsukuba, Venus, Teddy and Cones standard datasets. The evaluation concerns non occluded regions (nonocc), all regions (all) and depth-discontinuity regions (disc). In the first step of our algorithm, the WACD likelihood function is first performed on all the pixels. Applying the *winner-take-all* strategy, a label corresponding to the best estimated disparity is attributed to each pixel. The second step consists in selecting a subset of pixels according to their confidence

Table 1. Qualitative evaluation given by *Recall* and *Precision* measures.

(23)

(24)

different color spaces and different color constancy. The obtained results are also compared with those obtained from gray scale images. The algorithms are implemented in Visual Studio C++ 2008, using the OpenCV and IT++ libraries. The four datasets are given in Figure 5.

Fig. 5. The four datasets used for the evaluation (a) pontet: a level crossing in Lausanne–Switzerland (b) chamberonne: a level crossing in Lausanne–Switzerland (c) EPFL–Parking (d) pan: a level crossing in France.
