Research And Implementation Of Scene Adaptive Pedestrian Detection System

The universal pedestrian detection system is designed to provide universal pedestrian detection functions for users in a variety of application scenarios,such as video pedestrian detection monitoring in indoor surveillance scenarios,and assisted pedestrian labeling when labeling pedestrian data.However,existing pedestrian detection systems usually use a single pedestrian detection algorithm,thus cannot flexibly adapt to the complex and changeable usage scenarios in reality.For example,in the street pedestrian monitoring scenario,when the pedestrians are close and clearly visible,using complex and high-latency pedestrian detection algorithms has a relatively serious waste of computational resources;while when the pedestrians are far away and not easily distinguishable,using simple and low-latency algorithms may have more missed and wrong detection situations.To address the above problems,this paper proposes a scene adaptive pedestrian detection algorithm.This paper first proposes a simple and efficient scene adaptive scheduling strategy for the pedestrian detection algorithm.Specifically,scene classification is first performed on the input samples.We propose to classify the samples into simple scenes(larger pedestrian height)or complex scenes(smaller pedestrian height)based on the pedestrian height in the samples.For these two types of scenarios,we design high-performance and high-accuracy pedestrian detection algorithms respectively,and schedule the corresponding pedestrian detection algorithms according to the scene category of the samples in the prediction stage.Experimental results demonstrate the following advantages of this approach:1.Since samples with different pedestrian heights present large differences in image data,this approach makes the scenes well predictable,and at the same time,simple and complex scenes have different needs for pedestrian detection algorithms,which helps to quickly discover pedestrian detection algorithms compatible with the scenes;2.Pedestrian detection efficiency is significantly improved when using this strategy.The high performance pedestrian detection algorithm is more efficient in simple scenes,and the high precision pedestrian detection algorithm is more efficient in complex scenes.In addition,to further improve the accuracy of the pedestrian detection algorithm,this paper proposes a high-quality pedestrian sample selection algorithm that greatly reduces the annotation noise in the training set while avoiding as much as possible the differences in the distribution of training test data.The experimental results show that the accuracy of the pedestrian detection algorithm is significantly improved due to the improvement of the training set quality.Based on the scene adaptive pedestrian detection algorithm,a scene adaptive pedestrian detection system is designed and implemented in this project to help users monitor and obtain video pedestrian data.Detailed system test results show that the scene adaptive pedestrian detection system implemented in this topic can meet all the intended requirements.