*4.3. Examples of Selected Walking Cycles*

To alleviate the effect of occlusions and cluttered background, we propose to select the "best" walking cycle (a gait period). To further demonstrate the good performance of the proposed walking cycle selected method, we present two more examples in Figure 10 (we have given an example in Figure 5). All the candidate walking cycles are shown for each example, and the "best" walking cycle with highest score computed by Equation (3) is indicated by the red bounding box. From Figure 10a, we can find that the low portion of human in the first three images of the walking cycle *Frame#4-28* is partially occluded, and the walking cycles *Frame#29-52* and *Frame#40-66* both have cluttered background, while the walking cycle *Frame#14-38*, which is the selected one, has less noise than others. In Figure 10b, the walking cycle *Frame#3-25* suffers from occlusion, and the walking cycle *Frame#3-25* suffers from occlusion and clutter background, while the "best" one *Frame#43-67* has less noise. These examples demonstrate that our walking cycle method can select the ones with fewer occlusions and cluttered background. We believe that building a representation based on the selected walking cycle can alleviate the effect of occlusions and cluttered background.

**Figure 10.** Two examples of selected walking cycles on iLIDS-VID. The candidate walking cycles for each video is presented, and the "best" walking cycle we selected is indicated by the red bounding box. Note that the image sequence in the green bounding box in (**b**) is not a candidate walking cycle. The frame indexes are given on the left side of the corresponding image sequences.

Note that the image sequence *Frame#26-42* with the green bounding box in Figure 10b is not a candidate walking cycle. Since the low portion of a human in this sequence is heavily occluded, it is quite difficult to extract accurate motion information. While we can also observe that all the candidate walking cycles are complete gait periods. This validates the good performance of our motion information extraction method. Combining walking cycle selection and motion

information extraction, the proposed method obtains an accurate walking cycle with fewer occlusions and cluttered background, which leads to accurate temporal alignment and robust representation.
