Research On The Methods Of Pedestrian Detection Based On Combination Of Multi-Parts

Pedestrian detection is the foundation of further research like pedestrian recognition and behavior analysis. At present, the most popular method of pedestrian detection is to find a feature which describe the pedestrian effectively, and then use the methods of machine learning to train this feature of simple pictures, at last solve the problem of pedestrian detection by the methods of pattern classification.Currently most methods of pedestrian detection are unable to detect the occluded pedestrian, to solve this problem, the thesis uses a method to extract the features from multi-parts of pedestrian. Two pedestrian features models are established:one extracts the histogram of gradients features from the pictures of low resolution; one extracts an improved local binary pattern features of face, face and shoulder part, trunk part and legs from the pictures of high resolution. The result of experiments prove that the thesis improves the detected precision of occluded pedestrian. The main work of the thesis is as follows:1. Introducing the key technique of pedestrian detection. Pedestrian detection consists of extracting pedestrian features, using methods of machine learning to train for classifier of these features, and detecting the test pictures by the classifier. Nowadays, the most popular pedestrian features are wavelet features, histogram of gradients features, local binary pattern features and so on, and experiments show the pros and cons of these features. Classification Algorithms contain k-nearest neighbor, Bayesian, support vector machines, Adaboost algorithm and so on, the thesis analyzes and concludes the application of these algorithms in the pedestrian detection.2. The improved HOG and LBP features are applied to pedestrian detection. The thesis obtains the images of low revolution by down sampling from the simple pictures, extracts the HOG features and uses the principal component analysis to reduce the dimension of the features. Using support vector machines algorithm to train for classifier by the features of low dimension, reduce the training time and improve the rate of pedestrian detection. Combining the complex properties of the quaternion and LBP features gets an improved LBP features. The features retain information of three channels from color images which is robust to the color of background.3. Multi-instance learning and Bayesian net are used to training classifier. At first, the thesis extracts the improved LBP features from face, face and shoulder part, trunk part and legs in the images of high revolution, and uses the multi-instance learning to training classifier and detecting for every part of pedestrian. Then using the Bayesian net to construct a joint probability model, and combining the detected results from every part of pedestrian to training the classifier of pedestrian. The classifier can detect the occluded pedestrians and improve the accuracy of pedestrian detection.