A pedestrian comparison method based on multi-scale feature fusion

Abstract

The invention discloses a pedestrian comparison method for multi-scale feature fusion in the field of computer video processing technology. The invention combines various pedestrian features to enhance the stability and uniqueness of the comparison features in a multi-camera environment; at the same time, according to the characteristics in different Performance on the image scale, different pedestrian features will be compared on different scales: first compare and filter on a small scale, and then match the filtered pedestrians on a larger scale to ensure the comparison performance of each feature Under the premise of reducing the complexity of the method; and improving the existing texture features, a new comparison method based on salient feature points is adopted; this method introduces a semi-supervised distance learning method for distance function learning to reduce the traditional distance. The complexity of the learning algorithm training calibration process improves the matching accuracy.

Description

technical field [0001] The invention belongs to the technical field of computer video processing, and specifically relates to a comparison method for combining features of various pedestrians, first performing comparison and filtering on a small scale, and then matching target pedestrians on a larger scale. Background technique [0002] At present, pedestrian comparison technology is playing an increasingly important role in video surveillance, especially in the field of urban public security. Due to the similar characteristics of different pedestrians under the camera and the diversity of multi-camera network environments (view angle changes, illumination changes, mutual occlusion, etc.), pedestrian comparison under multi-camera faces severe challenges. Pedestrian analysis based on multi-camera has become a research hotspot in the field of computer vision. Research ideas generally include comparison algorithms based on statistical features, local feature points and distanc...

AI Technical Summary

Problems solved by technology

Pedestrian comparison method based on statistical features (see: Kviatkovsky, I.; Adam, A.; Rivlin, E., "Color Invariants for Person Reidentification," Pattern Analysis and Machine Intelligence, IEEE Transactionson, vol.35, no.7, pp.1622, 1634, July 2013) are usually relatively simple, stable, and low-complexity methods, and can achieve obvious results in simple scenarios. However, due to the use of histograms for feature statistics, local difference features are difficult. reflect

Pedestrian comparison method based on local feature points (see: C.Varytimidis, K.Rapantzikos, Y.Avrithis.WαSH: Weightedα-Shapes for Local Feature Detection. In Proceedings of European Conference on Computer Vision (ECCV 2012), Florence, Italy, October2012.) extracts the local feature points of pedestrians and compares them through the feature point matching algorithm. This method needs to match and calculate all the feature points or feature areas of pedestrians to obtain the similarity between pedestrians, so the complexity of the method Usually high, unable to meet real-time needs

The comparison method based on distance learning (see: Wei-Shi Zheng, Shaogang Gong, Tao Xiang. Reidentification by Relative Distance Comparison, PAMI2013, 2013, 35(3): 653-668) can compare The effect has been greatly improved, but the universality is not strong. Retraining is required for new scenarios. The training and calibration process is relatively complicated, and the space complexity of the method is high. It is still difficult to apply to the actual system at present.

Images