With our society rapidly aging, it is worried that the burden on families who have members in need of nursing care increase. Falling is one of the main factors that the elderly people have come to require nursing care. For prevention of falls in the elderly, gait trainings to enhance not only motor function but also cognitive function such as multi-target stepping task (MTST) to walk stepping on the assigned color target have been proposed. To evaluate the risk of falls, it is required to judge whether the subject steps on the correct target (target step judgment) and detect a behavior that the swinging leg crosses against supporting leg (cross step detection). In this study, to measure these indexes, a gait measurement system using a laser range sensor (LRS) is proposed. However, situations in which a leg is hidden from the sensor or the legs are close occur and lead to losing track of them or false tracking. To solve these problems, a novel legs detection method with five observed legs patterns and global nearest neighbor (GNN)-based data association with a variable gate based on the states of the legs are proposed. In addition, methods to judge target step and detect cross step based on the trajectory of the legs are proposed. From the experimental results of multi-target stepping task with the elderly, it is confirmed that the proposed system can improve the legs tracking performance, judge target step and detect cross step.