For UMIST datasets, we randomly selected 20% of the images from the database as training samples, with the remaining 80% used as test samples. Figure 3 shows some image samples from the UMIST dataset. We repeated 20 runs and report the average results and corresponding parameters in Table 2.
Fig. 3. Sample face images from UMIST database.
Table 2. Best average recognition rates of all methods on UMIST dataset.
Method K=1 K=3
PCA 85.37 ± 0.71(80) 85.37 ± 0.71(80) LPP 75.57 ± 0.96(80) 74.72 ± 1.59(80) MFA 83.55 ± 0.33(36) 81.95 ± 1.79(31) DNE 85.90 ± 1.74(76) 84.26 ± 1.64(77) DAG-DNE 87.72 ± 0.52(47) 87.23 ± 0.82(32) LBDAG_DNE 89.53 ± 0.18(16) 89.76 ± 0.46(19)
First, we consider the parameter selection. The nearest neighbor parameter K is selected from the set {1, 3}. Figure 4 illustrates the relationship between the accuracy and the value of 𝛽. From Fig. 4, we know that the accuracy is not the highest when 𝛽 =1, where 𝛽 is a tuning parameter that balances the tradeoff between intra-class infor‐
mation and inter-class information. The intra-class information and inter-class infor‐
mation play different roles in the classification task.
Figure 5(a) and (c) shows the accuracies of the four methods vs. the dimensionality of the subspace with different K. Figure 5(b) and (d) shows the relationship for the subspace dimension with the best accuracy. As seen in Fig. 5(a) and (c), the classification accuracies of all four algorithms increase rapidly. However, LBDAG-DNE has the fastest increase. From Fig. 5(b) and (d), we can see that LBDAG-DNE has the lowest discriminant subspace, which provides a good performance.
LBDAG-DNE: Locality Balanced Subspace Learning 207
(a)K =1
(b)K =3
0 1 2 3 4 5 6 7 8 9 10
65 70 75 80 85 90 95
Accuracy(%)
0 1 2 3 4 5 6 7 8 9 10
76 78 80 82 84 86 88 90
Accuracy(%)
Fig. 4. Average recognition rates vs.𝛽
Furthermore, Table 2 reports the best average recognition rates on the test sets for all of the methods, along with the corresponding dimension of the reduced subspace under different values of K. In spite of the variation in K, LBDAG-DNE has the highest recognition rate among these algorithms.
Based on the results of the handwriting and face recognition experiments, we can see that the classification performance of LBDAG-DNE is the best compared to DNE, MFA, and DAG-DNE. This suggests that the intra-class information and inter-class information have different degrees of importance for classification. In other words, they play different roles in the classification task. Moreover, the superiority of LBDAG-DNE was effectively demonstrated in all of the experiments. We could reduce the computa‐
tional complexity and improve the classification using LBDAG-DNE to extract the effective features.
208 C. Ding and Q. Sun
(a) K=1 (b) K=1
0 10 20 30 40 50 60 70 80
0.4 0.5 0.6 0.7 0.8 0.9 1
Dimensionality
classification accuracy
DNE MFA DAGDNE LBDAG-DNE
1 2 3 4
0 10 20 30 40 50 60 70 80
Algorithms
Dimensionality
DNE MFA DAG-DNE LBDAGDNE
(c) K=3 (d) K=3
0 10 20 30 40 50 60 70 80
0.4 0.5 0.6 0.7 0.8 0.9 1
Dimensionality
classification accuracy
DNE MFA DAGDNE LBDAG-DNE
1 2 3 4
0 10 20 30 40 50 60 70 80
Algorithms
Dimensionality
DNE MFA DAG-DNE LBDAGDNE
Fig. 5. Recognition rates for different parameters on UMIST database
5 Conclusion
The superior computing power of cloud computing makes it possible to utilize tuning parameters to select the best features. In this paper, we proposed a novel supervised discriminant subspace learning algorithm, called LBDAG-DNE, with the goal of learning a good embedded subspace from the original high-dimensional space for clas‐
sification. LBDAG-DNE maintains the intra-class and inter-class structure by constructing adjacency graphs and balances them by introducing a balance parameter.
More importantly, by introducing a balance parameter, it can also regulate the different levels of the intra-class information and inter-class information. Thus, LBDAG-DNE could find an optimal projection matrix. Experimental results show that LBDAG-DNE could achieve the best classification performance in comparison with several related state-of-the-art methods.
Acknowledgement. This work is supported by the National Science of Foundation of China, under grant No. 61571066 and grant No. 61472047.
LBDAG-DNE: Locality Balanced Subspace Learning 209
References
1. Roweis, S.T., Saul, L.K.: Nonlinear dimensionality reduction by locally linear embedding.
Science 290(5500), 2323–2326 (2000)
2. Tenenbaum, J., Silva, V.D., Langford, J.C.: A global geometric framework for nonlinear dimensionality reduction. Science 290(5500), 2319–2323 (2000)
3. He, X.F., Yan, S.C., Hu, Y.C., Niyogi, P.: Face recognition using laplacianfaces. IEEE Trans.
Pattern Anal. Mach. Intell. 27(3), 328–340 (2001)
4. He, X.F., Niyogi, P.: Locality preserving projections. In: Proceedings of Advances in Neural Information Processing Systems, pp. 153–160 (2003)
5. Xu, Y., Zhong, A.N., Yang, J., Zhang, D.: LPP solution schemes for use with face recognition.
Pattern Recogn. 43(12), 4165–4176 (2010)
6. Wang, S.G., Zhou, A., Hsu, C.H., Xiao, X.Y., Yang, F.C.: Provision of data-intensive services through energy-and QoS-aware virtual machine placement in national cloud data centers.
IEEE Trans. Emerg. Top. Comput. 4(2), 290–300 (2016)
7. Wang, S.G., Fan, C.Q., Hsu, C.H., Sun, Q.B., Yang, F.C.: A vertical handoff method via self- selection decision tree for internet of vehicles. IEEE Syst. J. 10(3), 1183–1192 (2016) 8. Fukunaga, K.: Introduction to Statistical Pattern Recognition. 2nd edn. Academic Press (2013) 9. Martinez, A.M., Kak, A.C.: PCA versus LDA. IEEE Trans. Pattern Analysis Mach. Intell.
23(2), 228–233 (2001)
10. Yu, H., Yang, J.: A direct LDA algorithm for high-dimensional data with application to face recognition. Pattern Recogn. 34(10), 2067–2070 (2001)
11. Yan, S.C., Xu, D., Zhang, B.Y., Zhang, H.J., Yang, Q., Lin, S.: Graph embedding and extensions: a general framework for dimensionality reduction. IEEE Trans. Pattern Analysis Mach. Intell. 29(1), 40–51 (2007)
12. Zhang, W., Xue, X.Y., Guo, Y.F.: Discriminant neighborhood embedding for classification.
Pattern Recogn. 39(11), 2240–2243 (2006)
13. Ding, C.T., Zhang, L.: Double adjacency graphs-based discriminant neighborhood embedding. Pattern Recogn. 48(5), 1734–1742 (2015)
14. Ding, C.T., Zhang, L., Lu, Y,P., He, S.P.: Similarity-balanced discriminant neighborhood embedding. In: Proceedings of 2014 International Joint Conference on Neural Networks, pp.
1213–1220 (2014)
15. Xu, J.L., Wang, S.G., Zhou, A., Yang, F.C.: Machine status prediction for dynamic and heterogeneous cloud environment. In: Proceedings of 2016 IEEE International Conference on Cluster Computing, pp. 136–137 (2016)
210 C. Ding and Q. Sun
Collaborative Communication in Multi-robot Surveillance Based on Indoor Radio Mapping
Yunlong Wu, Bo Zhang(B), Xiaodong Yi, and Yuhua Tang State Key Laboratory of High Performance Computing, College of Computer,
National University of Defense Technology, Changsha, China zhangbo10@nudt.edu.cn
Abstract. This paper considers a scenario where multiple sensing robots are deployed to monitor the indoor environments, and trans- mit the monitored data to the base station. In order to ensure favor- able surveillance quality, we aim at achieving a high throughput for the multi-robot system. We firstly establish the stochastic wireless channel model and derive the expression of the throughput. Then, we propose the non-collaborative and collaborative communication strategies, both adopting the joint frequency-rate adaptation based on the stochastic channel model. The experimental results have shown that the throughput can be largely improved with the collaboration between robots. Further- more, considering our surveillance scenario is approximate time-invariant (ATI), we propose the joint frequency-rate communication strategies based on proactive channel measurements, and the effectiveness of the strategies is validated by experimental results.
Keywords: Multi-robot collaborationãJoint frequency-rate communi- cation strategiesãRelaysãWireless channel modeling
1 Introduction
A multi-robot system aims at achieving challenging tasks or significantly improv- ing mission performance compared with a single robot, which demands consensus and cooperation among robots [1]. In this paper, we consider a scenario where a team of sensing robots are deployed to monitor an indoor area, and transmit the monitored data (e.g., videos) to a base station through wireless communications.
In the base station, the data will be analyzed for identifying abnormalities (e.g., an intruder) in the area.
For this multi-robot surveillance scenario, communication planning is demanded for maintaining reliable and high-throughput communications between each sensing robot and the base station. In [2], a rate-configurable robot was deployed to collect the generated data from several points of interest (POIs).
A multi-robot communication planning strategy was considered in [3], where the authors aimed at maximizing the connectivity probabilities of the multi-robot
c ICST Institute for Computer Sciences, Social Informatics and Telecommunications Engineering 2017 S. Wang and A. Zhou (Eds.): CollaborateCom 2016, LNICST 201, pp. 211–220, 2017.
DOI: 10.1007/978-3-319-59288-6 19
212 Y. Wu et al.
network by optimizing the routing variables. However, the communication strate- gies above just consider the case that the robots adopt the fixed communication frequency-allocation.
Against this background, in this paper we propose the joint frequency-rate communication planning and a series of strategies for multi-robot systems. The contributions of this paper are as follows:
In Sect.2, in order to support the joint frequency-rate communication plan- ning, we firstly establish the multi-frequency wireless channel model which reflects the received signal quality (RSQ) of each communication frequency at each spatial location. Then, the channel model is used to evaluate the communi- cation performance metric (e.g., throughput) and the packet error rate (PER).
The distribution of RSQ may be depicted as a stochastic model with three main parts: path loss, shadowing and multipath fading [4]. The multi-robot system can select the optimal frequency-rate setting to maximize the throughput according to the RSQ-location mappings.
In Sect.3, we propose two joint frequency-rate communication strategies based on the stochastic channel model. The non-collaborative communication strategy considers that each sensing robot communicates with the base station, and there is no information exchange among the sensing robots. Thecollabora- tive communication strategy allows the sensing robots to assist each other by relaying. For example, if a sensing robot is experiencing deep fading, relaying by other robots may greatly improve the throughput for supporting the monitored data transmission [5,6].
In Sect.4, we pinpoint the fact that the stochastic channel model may cap- ture the distribution of RSQ, however cannot exploit the exact RSQ at different locations for improving the throughput. Therefore, we propose to construct the RSQ-location mapping with proactive channel measurements, which may cap- ture the RSQ more precisely than the stochastic model. Especially, when the RSQ changes rapidly in spatial domain variation, while changing slowly over the time domain. The experimental results have proved that the actual indoor measurements can identify the approximate time-invariant (ATI) scenarios and a higher throughput may be achieved in comparison with the stochastic model- based strategies.
In Sect.5, the experimental results prove the effectiveness of joint frequency- rate communication planning, the collaboration strategy as well as the proactive channel measurements in improving the throughput of multi-robot systems.
2 Problem Formulation
We assume a N-sensing-robot system is deployed to monitor an area, where the surveillance route of each sensing robot is predefined and periodic. That is to say, when the sensing robot returns to the initial location, it may continue another loop. The sensors equipped on robots are responsible for collecting the environment information, and we may require the monitored data should be transmitted to the base station in real-time. Considering the high data sampling
Collaborative Communication 213 rates of the sensors on the robots (e.g., 4 Mbps per channel for a 1080P camera), we need to reasonably configure the communication settings, in the context of this paper, the frequency or the modulation and coding patterns, for the sake of maximizing the throughput of the multi-robot network based on the channel quality of the current location.
In order to avoid the interference between sensing robots, we adopt the fre- quency division multiple access (FDMA) mechanism, where the sensing robots are allocated orthogonal frequency channels for communications. However, we also allow dynamic frequency re-allocation during the operations. In an indoor environment, the geometric structure is often complicated which leads to the high complexity of the wireless communication channel. In order to get an opti- mal communication setting, we need to predict the received signal power at each spatial location and of multiple frequency bands.