Introduction

In recent years it has been observed a huge increase in the dementia related problems in the older population. A lot of new methodologies are being developed to recognize and observe the characteristics of people suffering from dementia. One such example is retracing the same path back and forth due to memory loss. The patient is observed walking, then stopping for a few moments or abruptly turning and then walking in the opposite direction following the same path.This discussion caters to recognition of such events using comparison between two consecutive walks. The comparison is done on the basis of position of the patient in each frame during the walk.

Methodology

We generate the xml1 and xml3 file by performing event recognition on SUP. SUP stands for Scene Understanding Platform which performs event and activity recognition.

In order to compute the relation between two consecutive walks that account for retracing of steps, we need to compare the positions at all points of the two walks. The information about spatial properties (2 and 3 dimensional features) are extracted from the xml1 file and accessed in python by converting the xml1 file to pandas data-frame. Similarly, the xml3 file is converted to pandas data-frame and all the frames corresponding to a walk are stored in a list.

Now, the next step is to compare every pair of frames (because they constitute a walk event) to the pair of frames prior to them.

For example g=[110, 119, 148, 178, 194, 208, 247, 295] here, 110-119 is one walk event,148-178 is another walk event and so on. So what we have to do is to compare the walk instance 147-178 to the walk instance 110-119 and walk instance 194-208 will be compared with 148-178. This comparison between the two walk instances is done on the basis of positions at each point of the walk. It has been taken into consideration to compare the positions of second walk in reverse order.

For example if frames 1325, 1350, 1367, 1397 are two walk events that correspond to retracing of footsteps, where 1325-1350 is one walk instance (going) and 1367-1397 is another walk instance (returning). We compare the positions of 1367-1397 to 1350-1325 in reverse order thus ensuring that the two walk events are a perfect candidate for step retracing.

These events are then checked for instances in which the person exits the scene and re-enters the scene following the same trajectory. Such events are of no interest from a clinical point of view hence these cases are dropped from the final list of frames that describe step retracing.

Further, to not miss the events of step retracing in a single walk,we propose to segment the walk and compare for positions using a sliding window. Each walk is compared for positions after the 20 frames of the walk. If the positions are closely related, the walk qualifies as an event of step retracing.

Tables

The output can be computed in the following format