Studies On Person Reidentification Algorithm

Intelligent video surveillance is an emerging research field in computer vision, and is commonly used in security, transportation, intelligent vehicles. Person re-identification is one of the core functions in intelligent video surveillance and has great academic value and application prospect. To achieve stable and accurate person re-identification in monitoring scenarios, this dissertation is dedicated to analyze person detection algorithm, low-level feature extraction algorithm and attribute-based re-identification algorithm. Based on these researches, a person re-identification system with decent accuracy is devised and the flowing work is done during this process.First, a person detection algorithm was devised and implemented based on motion detection and head-shoulder HOG feature detection. Using a background subtraction algorithm with multi-mode mean model, the foreground mask of surveillance image was generated, on which a relevance filter was posed and after-processing was done to extract the moving target. Afterwards, the head-shoulder HOG feature is extracted to avoid occlusion and then used to train a cascade Adaboost classifier. Later on, we devised a pyramid approach to search for people and then feed to the classifier and detect human existence.Then, based on the results of person detection, low-level features can be extracted. Color calibration is conducted using a dynamic threshold algorithm, and the person image is segmented to facilitate future calculations. Weighed histogram and color encoding is selected to be the primary color features. And Gabor filter along with Schmid filter is used for texture feature extraction.Finally, we use attribute based approach to achieve re-identification. Some human appearance attributes are carefully chosen according to several principles and training set are labelled. Low-level features of the training set are then fed to an SVM classifier to train a classification model which later on would be used to assign attributes to test instances. At last, feature fusion and weighing is conducted to merge attributes and low-level features to form a more comprehensive feature that will be used to calculate distance between the probe image and gallery set and achieve re-identification.The person re-identification algorithms presented by this paper were experimented on multiple datasets, shows that it has good accuracy and efficiency.