Polygonal approximation of a body shape

A convex set defines a region such as each segment of every pair of points is inside that region. A convex hull defines the minimal convex set containing a set of points.

We replace the coordinates of the body shape (i.e., the coordinates of the white pixels) by the ones from the convex hull of several sets. Each set of points belongs to the points of the body shape and contains nneighbor points.

• Step 1: LetSdenote the set of points of a body shape.∀x∈S, we find and draw the convex hull of V ⊂S, whereV is a subset ofScontainingxand itsn−1nearest neighbours in terms of distance along the body contour.

• Step 2: The step 1 produces several polygons ofnpoints. We only keep the lines on the border and we obtain the new shape.

The Figure5.8shows the different steps of the process. The green lines and points in the Figure5.8(h) define the original body shape, and the blue lines the shape approximated by convexity withn= 5.

The higher the valuenthe more convex is the shape (as illustrated in the Figure5.9).

The results demonstrated in the section5.3.4.1show the suppression of the gender when we apply the approximation by convexity. Indeed, the machine systematically interprets the final shape as female.

5.3.4 Experimental Results

In this part, we prove that our twode-genderizationmethods hamper gender detection like the inpaint-ing [3], JPEG scramblinpaint-ing with high level [6] of protection (i.e., encryption of DC and AC coefficients) and black masking methods. We explain these existing methods in the sections2.3and2.4. In the follow-ing, we conclude that the approximation using convexity (the second proposed approach) preserves the human activities contrary to the others.

5.3.4.1 Evaluation of gender detection

We evaluate the performance of gender detection using the same protocol described in the section 5.2.6.6, over the original videos, those generated by the first method (5.3.2) withα= 0,0.2,0.4,0.6,0.8,1, and those generated by the second method (5.3.3) withn= 3,5,10,20,30.

The Figures5.10and5.11show the accuracy of male and female detection for original (the black point), shape (the pink point) body images compared to the ones where the body contours are reshaped (the blue points). The random case (the red point) can be considered as the results of the inpainting, scram-bling and black masking methods. We can note that the performances decrease for the two approxi-mation methods. For example, whenα= 0.2using a male model (the point towards the top left corner

Chapter 5.Spatial-domain scrambling preserving the utility of visual surveillance 65

(a)

(b) (c) (d)

(e) (f) (g)

(h)

FIGURE5.8: (a) Original body shape, (b-g) step 1: drawing the convex hull of each point, x, and step 2:

its 4 neighbors (n=5), and (h) Keeping the lines on the border only.

in the Figure5.10), the rate of correct detection for females is about 10 % whereas the one for male is about 95 %.

The Figure5.10represents the results of the first method (5.3.2). We can observe that the closer to the model shape is the final shape, the more the tool classifies the gender of the final shape as the gender of the model shape. The Figures5.7(a)and5.7(b)show the selected reference models.

The Figure 5.11represents the results of the second method (5.3.3). We can observe that the higher the value ofnthe more the classifier tags the body images as female. This is mainly due to the convex approximation which seems to produce more female forms. Indeed, as it is shown in the Figure5.9and 5.14several classical women hairstyle appear, and in the lower body we can guess the presence of a skirt.

Chapter 5.Spatial-domain scrambling preserving the utility of visual surveillance 66

FIGURE 5.9: Original image and shape approximation using convexity with the associated average ac-curacy of gender recognition (in the second row) according to the value of the parametern(in the first

row).

FIGURE5.10: Results for the merging approach

The approximation using the convexity is a better approach than the other one. As with an intermediate value ofn(i.e., 10) the gender detection becomes weak whereas the motion of the arms and legs are much more visible than with the approximation using the merging. This can be observed in the video available online⁴. Moreover, for the first method, given the knowledge of the reference model andα, an attacker might easily reverse the transformation, and then retrieve the original shape.

FIGURE5.11: Results for the body approximation using convexity

4https://youtu.be/rpuIDLrHx3g

Chapter 5.Spatial-domain scrambling preserving the utility of visual surveillance 67 5.3.4.2 Evaluation of sport events classification

To evaluate the impact on event classification, we chose to test our algorithm on sports categorization.

We utilize Deepdetect⁵ to classify the sports and the UCF Sports [95] as dataset. We include further details of this tool and this database in the section2.5.4.

For each selected images, we apply inpainting, JPEG high-Level scrambling, black masking, shape detection (5.3.1) and approximations using convexity (5.3.3) withn= 3,5,10,20,30(Figure5.12).

FIGURE5.12: Respectively: Original image, shape, inpainting, scrambling, black masking and approxi-mated body withn= 10

The classification tool outputs an ordered list of classes from the best to the worst one. Therefore, we compute and show in the Figure5.13, the @k accuracy curve from k=1 to 10 (i.e., if the proper class is among the first k best results in the ordered list).

The results denote similar performances for sports classification in terms of shape and approximated body (∼10⁻²% of difference). Thus, removing the gender information by reshaping the shape does not impact the recognition of the sports compared when we use the original shape of the body. Moreover, the body approximation method achieves the best score compared to the black masking, the inpainting and the scrambling methods. However, the performances of sports classification5.13(a) for the black masking, inpainting or scrambling filters, are closer than expected from the original ones. This is due to the background which helps a lot in the recognition of some sports (e.g., acrobatic gym, horse riding, diving, golf or football). Indeed, in 5.13(b), where the tool recognizes sport from the RoI only, the performances drop more than 10 % compared to the original ones.

Thus, our approximation using convexity preserves the global movement of a person as well as people activities even in the absence of the background contrary to the other methods that in the best cases preserve only the global motion.