Object detection method and device based on classifier

Abstract

The invention discloses a target detection method and device based on a classifier. The method comprises the steps that an image pyramid is established through original images according to preset zoom factors; X image layers in the middle of the image pyramid are extracted, traversing sampling is conducted through a fixed sliding window, corresponding confidence degrees of sampled images are calculated through the classifier, and confidence sampling points corresponding to the Y sampled images with the highest confidence degree in each image layer are obtained; the confidence sampling points corresponding to the XY sampled images with the highest confidence degree are mapped to corresponding points in each image layer of the image pyramid, and each confidence sampling point and the corresponding point form a sampling chain; a window image of the corresponding image layer of all the points in each sampling chain is extracted through the sliding window, the corresponding confidence degrees of all the window images are calculated through the classifier, and then the window image with the highest confidence degree in each sampling chain is obtained; the window images with the highest confidence degree in all the sampling chains are mapped to detection result windows in the original images; the detection result windows in the original images are combined.

Description

technical field [0001] The invention relates to the technical field of intelligent video surveillance, in particular to a classifier-based object detection method and device. Background technique [0002] Classifier algorithm is a commonly used object detection algorithm in the field of intelligent video analysis. The most common classifiers, such as the Adaboost classifier, were proposed by Freud and Shapire et al. The Adaboost algorithm extracts knowledge of target concepts by learning labeled positive and negative samples, thereby generalizing to other unseen detection processes. [0003] In the traditional target detection process, it is roughly divided into two parts. One is to continuously scale the resolution of the image to form an image pyramid. Such as figure 1 As shown, the scaling ratio δ<1 is the scaling factor, layer 0 is the original image, and layer 1 and layer 2 are images obtained by scaling the original image once or twice... . The second is to use...

AI Technical Summary

Problems solved by technology

[0004] In summary, in the traditional technology, although the Adaboost classifier has a very short judgment time (microsecond level) for the sampled image, because it traverses and samples each layer, the total number of samples (usually reaching millions) is relatively large , so the time required for target detection is relatively long

In addition, since the Adaboost classifier performs binary judgment on the sampled images, the detection results are very dependent on the performance of the classifier. If the pre-trained Adaboost has high performance (for example, many positive and negative samples are used for training, etc.), the detection results will be relatively low. Accurate, but when the Adaboost classifier is not optimal, there will be too many false alarms in the detection results

Images