Scholarly article on topic 'Analysis of Socially Acceptable Smart Wheelchair Navigation Based on Head Cue Information'

Analysis of Socially Acceptable Smart Wheelchair Navigation Based on Head Cue Information Academic research paper on "Economics and business"

Share paper
Academic journal
Procedia Computer Science
OECD Field of science
{"Smart wheelchair" / "Head tracking" / "personal space."}

Abstract of research paper on Economics and business, author of scientific article — Razali Tomari, Yoshinori Kobayashi, Yoshinori Kuno

Abstract Smart wheelchair can be defined as a standard power electrical wheelchair that equipped with a mobile robotic technology to assist the user in a number of situations. Most of the smart wheelchair work focusing on safety issue and less work considers a socially acceptable issue. Since wheelchairs are normally used in human-shared environment, it is important to ensure the assistive motion generated from the wheelchair is safe and comfortable to the human in the surrounding. Here the framework for catering such an issue is proposed. The system initially infers human's state from head cue information. Next, the information is interpreted for modeling human's comfort zone (CZ) based on rules rooted from Proxemics concept. Finally, the wheelchair's motion is generated by avoiding both, the CZ and the in place obstacle. Experimental results demonstrate the feasibility of the proposed framework

Academic research paper on topic "Analysis of Socially Acceptable Smart Wheelchair Navigation Based on Head Cue Information"



Available online at


Procedia Computer Science 42 (2014) 198 - 205

Analysis of Socially Acceptable Smart Wheelchair Navigation Based on Head Cue Information

Razali Tomari a *, Yoshinori Kobayashib, Yoshinori Kunob

aFaculy of Electrical & Electronic Engineering, Universiti Tun Hussein Onn Malaysia, Parit Raja Batu Pahat 86400, Malaysia bGraduate School of Science & Engneering, Saitama University, 255 Shimo-okubo, Sakura-Ku Saitama 338-8570, Japan


Smart wheelchair can be defined as a standard power electrical wheelchair that equipped with a mobile robotic technology to assist the user in a number of situations. Most of the smart wheelchair work focusing on safety issue and less work considers a socially acceptable issue. Since wheelchairs are normally used in human-shared environment, it is important to ensure the assistive motion generated from the wheelchair is safe and comfortable to the human in the surrounding. Here the framework for catering such an issue is proposed. The system initially infers human's state from head cue information. Next, the information is interpreted for modeling human's comfort zone (CZ) based on rules rooted from Proxemics concept. Finally, the wheelchair's motion is generated by avoiding both, the CZ and the in place obstacle. Experimental results demonstrate the feasibility of the proposed framework

© 2014 The Authors.PublishedbyElsevierB.V. This is an open access article under the CC BY-NC-ND license (

Peer-review under responsibility of the Center for Humanoid Robots and Bio-Sensing (HuRoBs) Keywords: Smart wheelchair, Head tracking, personal space.

1. Introduction

In recent years, numerous methods have been introduced for developing smart wheelchairs to accommodate the need of disabled community. The comprehensive characterization and development trend can be found in [1]. One of the smart wheelchair components is the autonomous ability that accommodates more assistive features to users by helping them observe the surroundings and lead themselves to the destination safely. This component can be useful

* Corresponding author. Tel.: +607-453-7518; fax: +607-453-6060. E-mail

1877-0509 © 2014 The Authors. Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND license (http://creativecommons.Org/licenses/by-nc-nd/3.0/).

Peer-review under responsibility of the Center for Humanoid Robots and Bio-Sensing (HuRoBs) doi:10.1016/j.procs.2014.11.052

to users with cognitive or visual impairment, or those who fatigue easily [2]. Realizing the importance, there has been a great deal of research devoted in this area. Some recent results can be found in [3], which highlights the setups of various individual systems and their strategies used for navigation.

Since wheelchairs are normally used in human dominated environments (care centre, hospital and convenient store), to assimilate harmoniously it existence must be non-threatening and human must feel like the wheelchair is just another human walking around. Although it is impossible to change the wheelchair appearance to be exactly like human, we can tune its behavior to obey some of the human social norms. For example, in everyday human's interaction unwritten social norms dictate how human supposedly moving around each other; by teaching wheelchair the similar convention, we may expect the wheelchair perform socially acceptable movement that collision free, comfortable and natural to human. Research concerning this issue is normally known as social navigation planning, which is a subset of Human-Robot Interaction (HRI) in non-verbal and non-cognitive case.

HRI research has received much attention in recent years with the tremendous growth of sensing technology. Majority of HRI researcher focusing on direct interaction between human and robots such as face expression, or gesture recognition; apart from that the indirect interaction also play an important role especially for a mobile robot that must moving around people intensively such as robotic wheelchair and museum guide robot. A number of methods have been developed to let robot navigate around people naturally. One of the early ideas was proposed in [4] where they presented the components that may be required for developing human-friendly navigation systems. Sisbot et al. [5] investigated how a robot may approach a human by considering safety criteria, visibility criteria, and hidden zones through searching for the minimum cost path using A* algorithm. However due to the visibility to human requirement, the generated path may looks unnatural. C.-P. Lam et al. [6] proposed a system where robots and human can co-exist and navigate smoothly using six harmonic rules. In their system, several sensitivity fields of human and robot has been developed and utilized for planning the socially path. In [7] and [8], the researcher has demonstrated an adaptive system where a robot can navigate based on the information of a person seeking for interaction (PI). They determined the PI indicator based on the Case Based Reasoning (CBR) and then modeled the navigation system based on the person centered potential field. Y. Tamura et al. [9] proposed a method of predicting human's behavior during movement based on the information obtained from tracking human's legs. The robot predicted the human behavior, then deciding whether or not it should perform avoiding behavior by employing a social force model. R. Kirby et. al. [10] has extended standard personal space model for coping with tending to the right requirement when encounter human in an opposite case basis. This rule may applicable for handling a human in a wide area such as hallway and not really suitable for a small space like narrow corridor, e.g. in a corridor where human lies in right side of the robot, if the robot must pass through human's left space, no further movement can be suggested even though it is possible to navigate through human's right space. In [11] E. Pacchierotti et. al has investigated a minimum acceptable distance for letting robot passing a human in a narrow corridor. Based on the 10 subject's feedback, they found that the minimum comfortable lateral distance is 0.4 meter for a robot moving with 0.6 m/s, which is 0.05 meter less than the original proxemic's minimum personal space zone. Apart from distance, some researcher also concern about socially velocity as can be found in [12]. While this method begins to address ideas of planning speed around people, generally it does not directly consider social conventions.

Most of the mentioned human aware motion planners were developed for a service robot system where their objective is to serve humans and their highest priorities were given to humans. However, the situation is a little bit different for a robotic wheelchair in the sense that, wheelchair may have the highest priorities to move when humans are aware of their existence and only need to perform the socially avoiding action when such is not the case. For example, if humans are aware of the wheelchair in their path, they will usually stop or give a way for the wheelchair to move, thus wheelchair can maintain the original path. If wheelchair just treats all the obstacles in a same manner, there may occurred a situation where its path coincides with the human movement path and apparently bringing an awkward situation to them. To accommodate such requirement wheelchair should capable to estimate human intention (aware or not aware) and predict whether or not human will give a way. We need to consider such cases since there are situation where humans may be aware of the wheelchair but for a reason (elderly citizens or people with physical disability) cannot change their motion direction, and hence the wheelchair should keep away from them to avoid collision. This paper present extensive analysis of works accomplished from [13][14].

This paper is organized as follows. Section 2 presents the architecture of the proposed system with detail elaboration on each part. Next, Section 3 shows experimental results with discussion. Finally, the summary and future research is presented in Section 4.

2. System Overview

Extending a wheelchair planner to account for people requires the ability to identify and track human information (position and orientation) in the surrounding. Position can be defined as a projection of human's center mass on a ground plane, while orientation can be acquired in a different ways either from head poses, legs alignment, body orientation, and a direction of movement. In this project we opt to use head information for obtaining human's orientation. The head is the most visible sign of humans in occluded scenario and can be obtain from varying body posture. It also is useful to visualize human's perception field, that represents the direction where human's attention is oriented and simultaneously the awareness level [12].

The proposed method pipeline is outlined in Fig. 1. For each frame, the system starts by examine interesting objects in the wheelchair vicinity via x-z plane (z: depth direction). From this process, hypotheses of probable obstacle regions are generated; and by using blob analysis, non-standardize regions (with respect to humans dimension) were removed. The pre-validated regions are later on mapped onto x-y plane (RGB and Depth image), which provides information about most likely human's region of interest (ROI) distribution in both images. Since the pre-validation process is not enough, the (ROI) are further post-validated using upper human's silhouette (torso and head) cues. Only the survived areas are fed to the head detector for head localization and pose tracking. Later on, from the tracking attributes, personal space is modelled around human's spatial zone before executes the navigation task from the developed map. However, in this paper the focus is on analyzing the personal space model for the wheelchair usage by utilizing information from the head cue. For such a reason only detail elaboration about the head tracking module and personal space module is give in the following section.


Head Orientation Tracking

Personal Space Model

Wheelchair Motion

(Navigation Planner)

2D Map

Fig. 1. Method Pipeline

2.1. Head Orientation Tracking

In wheelchair coordinate system, the head's region at time t is defined as h = [xtyt wt at] where xt andyt is the region's centre coordinates, w, is the region's width and 0 is yaw head angle. To continuously predict such information, it must be tracked and hence the particle filter framework is selected based on work from [15]. To initialize a PF cycle, an estimation of initial target region is required. Here, initialization is triggered by an output from Adaboost based head detector with haar-like features [16]. However some pre-processing is needed since the detector is not an error free and we don't want to double initialize object that is currently under tracked. To satisfy the former requirement, we proposed depth assisted object segmentation using information from Kinect[17]. In short, the process initially examines interesting objects in the wheelchair's vicinity via x-z plane, and removing non-standardize region with respect to humans dimension via blob analysis. The pre-validated regions are then mapped onto x-y plane of RGB image and Depth image respectively for locating most likely human's region of interest (ROI) distribution in both images. The ROIs are further post-validated using human's silhouette information with HU moment features. Eventually, only the survived ROI's areas are fed to the head detector node. By implementing such process we manage to reduce the amount of spurious head detection and the processing time.

From the previous process, we have a numbers of head's localities in hand (hd). To assimilate such information into the tracker, a cross check procedure is performed between hd and currently tracked head regions (hc) using Euclidean distance measure. hd are assigned as newly tracked objects if its distances between all hc regions exceed the minimum threshold. Beyond that, we assume that the hd is currently under tracked. For each hc, PF samples at time t are projected according to equations (1) where n is a number of particles, Of is 8 points optical flow distributed evenly within hc region, En is the wheelchair's ego-motion obtained from IMU sensor, and the random vectors of provide the system with a diversity of hypotheses.

h= h+ Of" + E" + N" (1)

Observation model is used to evaluate particles confidence level by computing its weight. We use two evaluation methods based on image contour, and seven Adaboost cascades for classifying frontal, left and right face respectively. Overall weight of each particle is computed by fusing the likelihoods from the image contour and the cascade classifier. Eventually, current state of each target h/, is estimated by using the average weight of all PF samples. Based on the htc, spatial placements of heads location in world coordinate hw= (xw' yw zw), is recovered using depth data. If hw does not lies within the valid blobs region, then missed tracking is assumed occur and target tracks is deleted from the scene. This procedure greatly reduces the amount of spurious tracking and simultaneously reduces the risk of accumulated tracking error. The hw state provides a plausible human's location in x-z plane including estimation of head yaw angle, and is used to encode human's PS model. Fig. 2 shows example of the detected head tracking information in indoor environment.

Fig. 2. Sample of Head detection with pose tracking result

2.2. Personal Space Model

In indoor environment where space is a constraint, we need to develop the human personal space model using a minimum buffer without sacrificing the social effect. Given a human head location viewing from the obstacle map (xw,zw), we define the personal space around human using combination of two semi-ellipses. The model can be parameterized by equation (2) and where xw and zw is the center of both ellipses, a is the yaw head angle, a and b are the parameters that control the minor and major axis respectively.

For the first semi-ellipse (from 180 -360 °) the minor and major axis are defined by a, which is a minimum human's personal radius (we set the value to be 0.4 m based on the study from [11]). As for the second ellipse (from 0 -180 °), the major axis b is twice of a, i.e. 0.8, to satisfy the rule of human is more protecting their frontal space as oppose to the back side. Example of the generated PS shape is shown in Fig. 3. It can be seen that the major axis of the personal space is driven directly by the head orientation angle.

x(ff) = x + a cos(0)*cos(-a) - b *sin(0)*sin(-a) fb' if 0°<8< 180°

m\ w ■ , x ,„ ■ „ , x ' \a if 180° <9< 360° (2)

z(ff) = z + a cos(0)*sin(-a) + b *sin(0)* cos(-a) ^ J

3. Result and Analysis

To analyze the performance of the proposed method, we performed three types of experiments as follow: 1) Evaluating the area corresponds to the human's frontal space. 2) Evaluation of PS model in two different encountering situations. 3) Evaluation of wheelchair's movement in multiple human's coordination.

3.1. Frontal space (FS) relation with personal space (PS)

In the first experiment, we interested to evaluate the rule of "humans are more protecting on their frontal space (FS) as oppose to the backside" for modelling the PS by using head cue information. For modelling the PS using HP, there is an abstraction on how to define the human's FS. For example, when human's BO does not coincide with HP, it is unclear which zone can be considered as the FS, either in front of the body or in front of the face. To rectify such a problem, we conduct an experiment to approximate the human's FS when BO and HP is in differ form. The experiment is performed in real human scenarios where a participant sits on a bench in a main hallway and was asked to performing two gazing actions when a people pass through: (1) Gazing straight (GS), i.e., BO coincides with HP (2) Gazing at the approaching person (GP), i.e., BO differs with HP.

(a)i (a)ii (b)i (b)ii

Fig. 4. Sample snapshots during evaluating wheelchair actions for avoiding a human using SP Without participant on the bench

Fig. 4 shows some of the sample images taken during the experiment when there is no participant (Fig. 4 (a)i and (a)ii) and when the participant seat on the bench ( Fig. 4 (b)i and (b)ii). It can be seen that for the first situation, people passing through the bench with a very near distance (Fig.4 (a) i and (a) ii). However, when there is a participant sitting on the bench, the situations change drastically. People tend to create more distance when passing through the bench. This initial result clearly showed that, people passing through a human in quite a distant formation as oppose to a non-human. Such behaviour should be replicated into a wheelchair platform to ensure a socially acceptable motion can be generated.

100 200 300 400 500 600

50 100 150 200

250 ■■;■."■"' 300 ■■

Pedestrian flow

50 100 150 200 250 300 350

100 200 300 400 500 600 Pedestrian flow

(a) (b) (c)

Fig. 5. Person trajectories when participant initiate: (a) GS action. (b)(c) GP action.

To further analyze the effect of the generated distance with respect to the gazing action, we placed a laser sensor near the participants for measuring the relative pedestrian distance, and recorded a video (40 minutes) for annotating the results of the study. Our hypothesis is that, the zone where the participant gazing at will exhibit larger distance

separation from pedestrian compared to other zone. Fig. 5 shows the pedestrian's trajectories when the participant executing the GS and GP actions. The participant centre location (PCL) is at (340 cm, 280cm) and relative distance is measured from pedestrian location to the PCL through x axis (RDx) or y axis (RDy). Based on this figure and from video annotation, we found that when the participant performing the GS action (Fig. 5(a)), if the pedestrian is approaching at RDy smaller than 30cm, they will change their route by making the RDy larger starting at around RDx = 110 cm, maximum at PCL and reducing the distance later as they moving far from the participant. When the participant performing the GP (Fig. 5 (b) and (c)), most of the time, pedestrian will approach at maximum RDy at RDx > 200 cm, and as they moving forward, the RDy reduce gradually. From this finding we draw a few conclusions. 1) When the BO coincides with the HP, the separation distance is largest in front of human's area. 2) When the BO differs from the HP, the separation distance is maximal in human's gazing zone, and BO give less effect to the separation distance. Therefore we conclude the PS size in where human's gazing at, should have wider separation distance as oppose to the non-gazing zone

3.2. Personal space (PS) relation with wheelchair movement

The PS model shape will evolve according to the human's gaze, in a sense where the major axis b will always in front of human head. For evaluating the feasibility of this model, we conduct an experiment using 10 participants in which the wheelchair avoids each participant in static/walking condition, by using the PS model and without using the model. We call the former wheelchair's motion as social path (SP) and the latter as conventional path (CP). The wheelchair encounters participants (static/walking) from the opposite and perpendicular directions. Sample snapshots during the experiment with trajectories illustration are depicted in Fig. 6. It shows (a) Opposite-static case, (b) Opposite-moving case, (c) Perpendicular-static case, and (d) Perpendicular-moving case.

(a) (b) (c) (d)

Fig. 6. Sample snapshots during evaluating wheelchair actions for avoiding a human using SP and CP: (a) Opposite-static case (b) Opposite-moving case (c) Perpendicular-static case (d) Perpendicular- moving case

The participants were asked to evaluate the wheelchair motion when performing the SP and the CP. Each participant evaluated the motion in terms of how comfortable and natural in a scale of 1 to 5: "definitely no", "very small effect", "small effect", "large effect", and "definitely yes". Fig. 7(a) shows the average participants' responses to the questionnaire. We conducted a repeated measures analysis of variances (ANOVA) for the single target setting. It shows no significant differences in case (a) (F [1, 19] =0.012, p =0.915) and case (b) (F [1, 19] =0.2975, p =0.5987). However, significant differences are found in case (c) (F [1, 19] =16.57, p =0.0028) and case (d) (F [1, 19] =18.45, p =0.002). Participants rated CP and SP with average scores of 2.7 and 4.5 respectively in case (c), while in case (d) the average scores are 2.7 and 4.4 respectively.

From this result we may conclude that in opposite cases, the SP can improve only small amount of humans' comfortable level, while in perpendicular encountering cases, the SP can significantly outperform the CP in terms of human-friendly navigation. We can interpret this result from human's awareness level. In opposite cases humans are aware of the wheelchair existence, therefore it does not affect their comfortableness much whether the wheelchair avoids with or without using the PS. In contrast, in perpendicular cases, humans may not be aware of the wheelchair, hence the SP is more preferable and comfortable compared to the CP.

Based on this finding, we have modified our PS model to consider the human awareness issue. The human awareness can be measured from his/her field of view (FOV) as illustrated in Fig. 7(b). Although human's FOV is approximately 180° , the effective zone (EZ) is around 120° . If the wheelchair encounters a human in his/her EZ

(positions 3/4/5), s/he is almost certain to notice it. In such cases, even though the wheelchair approaches him/her a bit closer, s/he still feels comfortable. Based on this finding, we modified the PS model so that b in equation (2) reduce to 0.4m when in the EZ situation and remain 0.8m in non-EZ situation. Sample of the generated PS that consider aware and unaware cases are shown Fig. 7 (c). When performing the navigation task, the wheelchair should respect the PS and therefore it will normally react early when seeing humans. If the wheelchair refrains from entering the PS zone, it is performing safe and comfortable motion. Without considering the PS, the wheelchair exhibits only safe motion.

o ^ .1 ? i 1 J l T3 I ÏI

Non-EZ = 240°\ _

>'' EZ = 120 ° M

(a) (b) (c)

Fig. 7. (a) Human evaluation of wheelchair movement for four trails of SP and CP. (b) Human field of view (FOV) zone illustration. (c)

Visualization of the PS shape (denoted by red marker)

3.3. Personal space (PS) evaluation in multiple human coordination

The final experiment was designed to analyze the wheelchair's motion for avoiding multiple humans (static/moving) in which their position and coordination were pre-arranged as follows: (1) Two people standing and facing each other. (2) Two people walking side-by-side and moving towards the wheelchair, (3) One person moving towards the wheelchair while another person moving in the perpendicular direction. We compared the movements when the wheelchair considered the PS and did not. The former motions are labeled as social path (SP) while the latter are marked as conventional path (CP). Experimental setup and motion results for scenarios (1) to (3) are listed in Figs.8 (a) to (c).

Fig.8. Multiple humans' coordination tested by the system.

Figs.8 (a) and (b) show that the CPs might be the shortest wheelchair's paths in these cases. However these paths caused the wheelchair to pass through the human frontal space in close proximity in scenario (1), and in the middle of two people in scenario (2). In scenario (3), the wheelchair passed through in front of the person coming from the right in the CP, whereas it moved through behind the person in the SP. Although the generated paths might not be shortest, the SPs did not interfere with the PSs of the people and kept comfortable proximity to them. To sum up, the experimental results demonstrate that the proposed system can generate a collision free path that does not invade people's PSs, thereby can provide safe and comfortable interaction with people around.

4. Conclusion

In this paper, an analysis of framework for socially acceptable wheelchair navigation in human-shared environments based on head cue information is proposed. From the first experiment, it is conclude that the frontal space of the PS model is evolved based on head orientation changing instead of body orientation changing. Next, from the second experiment we found that human awareness does statistically give significant effect to the PS shape, in a sense that only minimum PS is required when human is aware and maximum PS is needed when human is not aware. In the final experiment, we have showed that by using the PS the generated path is less interfere to the human and from social perspective is comfortable to human.

In future work, we intend to consider wheelchair's user comfortable issues as one of the planner constraint. Since autonomous wheelchairs transport humans, it is important that they also consider the persons that they carry.


The authors would like to thank to Ministry of Education (MOE) and Universiti Tun Hussein onn Malaysia (UTHM) for supporting this research under Research Acculturation Grant Scheme (Vot. no. R033).


1. R. C. Simpson, Smart wheelchairs: A literature review, Journal of Rehabilitation Research and Development, 2005, 42(4), pp. 423-436.

2. R. C. Simpson, E. F. LoPresti, and R. A. Cooper, How many people would benefit from a smart wheelchair?, Journal of Rehabilitation Research and Development,2008, 45(1), pp. 53-72.

3. R. Grasse, Y. Morere and A. Pruski, Assisted Navigation for Person with Reduced Mobility: Path Recognition through Particle Filtering (Condensation Algorithm), Journal of Intelligent Robot System, 2010, 60, pp. 19-57

4. R. Alami, I. Belousov, S. Fleury, M. Herb, F. Ingrand, J. Minguez, and B. Morisset, Diligent: Towards a Human-Friendly Navigation System, In Proc. of IR0S'02, 2002, pp. 21-26

5. E.A. Sisbot, L. F. Marin-Urias, R. Alami, and T. Simeon, A Human Aware Mobile Robot Motion Planner, IEEE Transaction on Robotics, vol. 23(5), 2007, pp. 874-883

6. C.-P. Lam, C.-T. Chou, K.-H. Chiang and L.-C. Fu, Human-Centered Robot Navigation-Towards a Harmoniously Human-Robot Coexisting Environment, IEEE Transaction on Robotics, (99), 2010, pp. 1-14

7. S. Tranberg, M. Svenstrup, H.J. Anderson and T. Bak, Adaptive Human-Aware Navigation based on Motion Pattern Analysis, In. Prof. of IEEE International Symposium on Robot and Human Interactive Communication, 2009, pp.927-932.

8. M. Svenstrup, S.T. Hansen, H.J. Andersen, and T. Bak, Adaptive Human-Aware Robot Navigation in Close Proximity to Humans, Intl. J. on Adv. Robotic Systems, 2011, 8(2), pp. 7-21 .

9. Y. Tamura, T. Fukuzawa and H. Asama, Smooth Collision avoidance in human-robot coexisting environment, In Proc. of IR0S'10, 2010, pp. 3887-3892.

10. R. Kirby, R. Simmons, and, J. Forlizzi, COMPANION: A Constraint-optimizing method for person-acceptable navigation. In. Proc. IEEE Intl. Sym. on RO-MAN09, 2009, pp. 607-612.

11. E. Pacchierotti, H.I. Christensen, and P. Jensfelt, Evaluation of Passing Distance for Social Robots, in Proc. 15th IEEE Intl. Symp. on RO-MAN06, Hatfield UK, 2006, pp. 315-320.

12. D. Shi, E.G. Collins Jr., A. Donate, X. Liu, B. Goldiez, and D. Dunlap, Human-Aware Robot Motion planning with Velocity Constraints. In Proc. Intl. Symp. On CTS08,2008, pp. 490-497.

13. R. Tomari, , Y. Kobayashi, and Y. Kuno, Enhancing Wheelchair Maneuverability for Severe Impairment User", Intl. Journal of Advanced Robotic Systems, 2013, 10, pp. 1-13

14. R. Tomari, Y. Kobayashi, Y. Kuno, "Socially Acceptable Smart Wheelchair Navigation from Head Orientation Observation", International Journal on Smart Sensing and Intelligent System, 2014, 7(2), pp. 630 - 643

15. Y. Kobayashi, D. Sugimura, Y. Sato, H. Hisawa, N. Suzuki, H. Kage and A. Sugimoto,3D Head Tracking using The Particle Filter with Cascade Classifiers, In Proc. British Machine Vision Conference,2006, pp. 37-46.

16. P. Viola and M. Jones, Rapid object detection using a boosted cascade of simple features, In Proc. of International Conference on Computer Vision and Pattern Recognition, 2001, pp. 511-518.

17. R. Tomari, Y. Kobayashi,and Y. Kuno, Multi-View Head Detection and Tracking with Long Range Capability for Social navigation Planning, In. Proc. of ISVC'11, 2011, pp. 418-427.