Pedestrian detection method and system of combination of deep learning network and artificial characteristic

Abstract

The invention provides the pedestrian detection method of the combination of a deep learning network and an artificial characteristic. The method comprises a deep learning network training step and adeep learning network detection step. The invention also provides the pedestrian detection system of the combination of the deep learning network and the artificial characteristic. In the invention, multithreading and parallel calculation are used to accelerate an artificial detection method, calculating time is obviously shortened, the mode of combining the deep learning network and the artificial characteristic is adopted, accuracy is high, a false drop rate is low and robustness is good.

Description

technical field [0001] The invention relates to the technical field of pedestrian detection, in particular to a pedestrian detection method and system combining a deep learning network and artificial features. Background technique [0002] Pedestrian detection is the process of detecting pedestrians from visual materials, mainly pictures and videos. Pedestrian detection is a very basic task in computer vision. On this basis, many important topics such as pedestrian recognition and pedestrian tracking can be derived, such as security and behavior analysis. [0003] With the rapid development of computer vision algorithms and computer hardware, the accuracy of pedestrian detection has increased. Since the introduction of deep learning into computer vision, its accuracy has been further improved. Before deep learning, people mainly artificially manufactured and extracted features according to the characteristics of image color and texture in specific scenes, and then completed...

AI Technical Summary

Problems solved by technology

However, although this patent is based on the deep learning network for pedestrian detection, it first needs to preset and learn the optimal zoom ratios of different sub-regions, and then scale the images of different sub-regions according to the corresponding zoom ratios and input them into the deep learning network for detection , the steps are relatively cumbersome, and in order to reduce the false detection rate, it needs to determine the correspondence between the pedestrian's position and size, the purpose is poor and the effect is open to question

[0005] To sum up, there are two disadvantages in pedestrian detection with artificial features. First, the production of artificial features needs to be specially designed according to the color, texture and many other information of the image. The balance between the detection effect and the consumption of computing resources cannot be better balanced; the second is that artificial features are relatively weak in the semantic understanding of images.

With the help of the GPU, the deep learning network has greatly improved the effectiveness, robustness, and semantic understanding of feature extraction. However, the existing pedestrian detection based solely on the deep learning network has low detection accuracy. High, poor system robustness and other issues

[0006] In conclusion, a single method has limitations and cannot achieve fully satisfactory results

Images