Implementation And Optimization For Part-based Human Body Detection

IVS Intelligent Video Analysis (Intelligent Video Ssytem) technology which is processing digital image and video with pattern recognition algorithms for visual image analysis. By separated the background and object of the scene, the targect object which appeared in the camera scene can be analyzed and tracked. The key to IVS is the target detection algorithm, object detection algorithm is the key to realize the intelligent video technology, is a critical step of IVS. Improving the performance of object detection algorithm will greatly improving the performance of intelligent video systems. Object detection algorithm can also be used alone to detect and count humans and vehicles in a carema scene, the importance and prospects of object detection is gone without saying.Part based object detection system which from P. Felzenszwalb et al is based on mixtures of multiscale deformable part models. This system is able to deal with highly variabled object classes and achieves state-of-the-art results in the PASCAL object detection challenges. This part based system relies on new methods for discriminative training with partially labeled data. A latent SVM is a reformulation of MI-SVM in terms of latent variables. Sloving semi-convex problem, constructing a stochastic gradient descent algorithm to train the model parameters and handle a large number of samples by using data mining algorithms, the algorithm of part based object detection is built up.The performance of the detector design in this paper is beyond the detection performance of latest model of P. Felzenszwalb et al, detection speed is also much improved, it can be achieved in near real time (ie, 200ms / frame). In this paper, in order to obtain the real result and do analyze of the result, we buildup a real scene datasets for testing the training and detection algorithm implementated in C programming language. Through the algorithm optimization, speed optimization and optimized for the actual scene, our detection algorithm is beyond the standard detection algorithm which based on parts base.This paper done the optimization of the algorithm in depth, completed the task of the expected, achieve the expected performance and speed requirements, also meet further constrain for the DSP platform migration.