Person Re-identification Based On Matching-CNN

With the development of the construction of sapiential city, information technology is play-ing an increasingly important role. Intelligent video surveillance system as an important part of the security system in the sapiential city, is becoming more and more concerned by the public. As an important part of the intelligent video surveillance system, the person re-identification technology has gradually aroused the interest of researchers. However, due to the complexity of the problem itself, the current solution is still very immature. There are a lot of difficulties in person re-identification, for example, the shooting angle is not unified, and the light condition is poor, the environment and the pedestrian posture is various. How to effectively solve these problems and improve the performance of person re-identification, promote the algorithm to the application of real life still need a lot of effort.In this paper, we propose three improved schemes for person re-identification based on matching-CNN. Aiming at the problem that the common features are difficult to distin-guish between the matched pair and non-matched pair, this paper proposes a new method to improve the matching rate of person re-identification by feature enhancement. For the cur-rent research program have failed to take full advantage of low-level visual features of the pedestrian image, this paper presents a method to effectively combine the HSV color feature and the texture feature of LBP after the use of feature enhancements. Since the amount of data on the open database currently used by the researchers are quite small, and manual pro-duction, marking the large scale of the special database is a very arduous process, a method for further adjustment of the model is extracted by using a large amount of unlabeled and unstructured network video, which effectively utilizes the hidden information contained in a large amount of data resources. The three parts of this paper are as follows:1. By introducing the idea of feature enhancement to the traditional network model and use triples as the training sample, then do pixel-level superimposed on the low-level feature maps which is a good representation of the low-level features extracted by convolution neu-ral networks, a feature enhancement is performed on matched pairs and mismatched pairs. In continuous iterative learning, the energy of the matched pair's feature vector is more con-centrated, but the energy of the non-matched pair's feature vector is more dispersed, thus widening the metric distance of matched pair and non-matched pair in the feature space. Feature enhancement can filter out irrelevant details while retaining the key features needed for pedestrian recognition.2. The color feature and texture feature are two key low-level visual features of the image, general network just use the RGB color features, not fully tapped the original image infor-mation. By simultaneous inputting the HSV color feature map and the LBP texture feature map, which obtained from the RGB color features of original sample, to the convolutional neural network, and combine the two key features of high level representation in the net-work using machine learning, the hybrid feature vector which contains the color feature and texture feature was obtained, which can improve the performance of the algorithm.3. Because of the difficulty of the training data acquisition, it is often the research personnel to design a specific scene, and then through a large number of boring and heavy manual selection to get the sample data. And the sample size of the open database in this area is very limited. From the massive video data, a large number of pedestrian samples are extracted. The obtained samples were roughly automatically classify and organized. Then, the data is used to modify the trained model, and the information of the massive video data is excavated, which makes the algorithm more robust, robust and universal.