A real-time pedestrian legs detection and tracking system used for autonomous mobile robots

In this paper, we develop a real-time pedestrian legs detection and tracking system that emphasizes on targeting the lower part of a human body. In the legs detection procedure, we evaluate two kinds of classifiers based on multilayer perceptron (MLP) and support vector machines (SVM) to determine whether pedestrian legs appear in a frame of the video sequence captured by a single webcam equipped with an autonomous mobile robot. In the legs tracking procedure, we extract both the color and edge information of the region of detected pedestrian legs as features, and then apply a particle filter to tracking them instantly. The experimental results manifest that our developed system is able to detect multiple pedestrian legs rapidly, and has high detection rates in both the indoor and outdoor cluttered backgrounds. The SVM-based classifier performs better than the MLP-based one does. Using the former classifier, the detection rate has more than 95.0% in indoor environments and is over 92.7% in outdoor environments. Furthermore, eight kinds of pedestrian walking orientations are also tested, and the average detection rate is over 90.0%. Additionally, the tracking rates of a video sequence containing a single pedestrian leg and multiple pedestrian legs are 97.0% and 90.0%, respectively, and the tracking performance attains 9 to 12 frames per second.