Pedestrian Detection Based On Uncertainty Theory And Machine Learning

Recently, with the rapid increase of the amount of vehicles, traffic environment is getting much worse. The situation of traffic safety on urban roads is becoming a critical issue. Pedestrians are in the most disadvantaged position in the traffic system, and also they are the most vulnerable groups. Data shows that pedestrian casualties in traffic accidents are very high. Therefore, developing an effective pedestrian detection algorithm is an effective measure to reduce the number of traffic accidents, enhance pedestrian safety and improve the efficiency of the city. This paper aims at the application of actual engineering, and designs a pedestrian detection based on uncertainty theory and machine learning.In view of the pedestrian appearance, complex scenes, illumination changes and other factors, it is difficult to detect pedestrian from the complex traffic scenes based on a single or few characteristics. Dempster-Shafer theory is a very efficient fusion theory based on uncertainty theory, which is a good way to detect pedestrian. Firstly, according to traffic psychology and behavior character of pedestrian, extract information, such as shape and motion, then use monte carlo method to get detection ability of characteristic of pedestrian detection, In next step, implement the data fusion based on the D-S evidence theory, and adjust the fusion process until the detection accuracy finally meets the supposed target..Integrating the advantage of neural network and the features of pedestrians which are described in the previous chapter, this article proposes a method of pedestrian detection based on Back Propagation neural network. At first, the six features of pedestrians: length-to-width ratio, width, area, speed, trajectory smoothness degree and motion vector field are as the input for BP network. Then analyze the hidden layer, hidden layer node and effective function to create an effective neural network construction. The output of the network is the detection result. Through these analysis and comparison, it’s found that the six features of pedestrians built upon each other, and can get the good measuring result. BP network shows its value of application and feasibility in pedestrian detection.The traditional deep learning should meet a basic assumption: the training data and test data are drawn from the same distribution and the same feature space. However, such assumption is untenable generally. When the application conditions change, we must collect training set-training-model again. So the deep learning and Support Vector Machine are used in pedestrian detection to automatically build a classifier by learning, which from a set of previously classified documents, the characteristics of categories. Experimental results show that the algorithm has good detection effect and it can detect the pedestrian quickly when the traffic conditions changes.