دانلود رایگان مقاله رویکردهای مختلف برای سیستم بازشناسی هویت

Abstract

An ensemble of approaches for reliable person re-identification is proposed in this paper. The proposed ensemble is built combining widely used person re-identification systems using different color spaces and some variants of state-of-the-art approaches that are proposed in this paper. Different descriptors are tested, and both texture and color features are extracted from the images; then the different descriptors are compared using different distance measures (e.g., the Euclidean distance, angle, and the Jeffrey distance). To improve performance, a method based on skeleton detection, extracted from the depth map, is also applied when the depth map is available. The proposed ensemble is validated on three widely used datasets (CAVIAR4REID, IAS, and VIPeR), keeping the same parameter set of each approach constant across all tests to avoid overfitting and to demonstrate that the proposed system can be considered a general-purpose person re-identification system. Our experimental results show that the proposed system offers significant improvements over baseline approaches. 

5. Conclusion

In this paper we run experiments to develop an ensemble of person re-identification systems that works well on different datasets without any ad-hoc dataset tuning. Therefore, we are quite sure that our approach is stable and could be used in different image conditions. For improving the state-of-the-art approaches, different color spaces, texture, and color features for describing the images were explored. We also considered different distances for comparing descriptors. Among the tested distances, the best performance was obtained with the Jeffrey Divergence measure. The new methods proposed in this paper were tested across several benchmark databases: CAVIAR4REID; VIPeR; VIPeR45; IAS. The experimental results demonstrate that the proposed approach provides significant improvements over baseline algorithms. The VIPER45 is a new dataset of 45 image pairs taken from VIPeR that focus on difficult samples with strong pose changes and with subjects wearing similar clothing. It was created because human beings were tested in [7] in a dataset that was built in a similar fashion (i.e., using 45 difficult image pairs extracted from VIPeR). It is thus possible for other researchers in person re-identification to use VIPeR45 for approximately comparing the performance of their computer vision systems with the performance of human beings. A drawback of our approach is computational time, which is not real-time, i.e., using MATLAB code. However, several methods used in our approach are internally highly parallelizable. The main focus of this paper was not on computational speed; our goal was to produce an approach that could match human performance. Unfortunately our results show that this goal has not been achieved (our ensemble obtains a Rank(10) of 65%, while a human being obtains 100%). Nonetheless, we have succeeded in producing a stable general-purpose Table 16 Computation time in seconds. FUS1 FUS2 FUS3 NogBiCov 1.55 8.15 10.25 4.82 Table 17 Comparison with the Literature. VIPeR CAVIAR4REID Here 22.9 56.2 25.3 65.0 OR_CPS 21.84 57.21 9 47 OR_SDALF 19.87 49.37 – eBicov 24.34 58.48 – OR_CI 24.00 58.00 9 45 kBiCov 31.11 70.71 – MCC [1] 15.19 57.59 – KISSME [11] 19.60 62.60 – PCCA-rbf [2] 19.27 64.91 – 152 L. Nanni et al. person re-identification system that offers significant improvements over baseline approaches.