Research Of Pedestrian Detection Based On Visconti2 7502

Aiming at the frequent traffic accidents in traditional driving field,people's demand for intelligent driving and safe driving is more and more urgent.Pedestrian detection as one of the key search in the intelligent driving,has become a hot research topic in machine learning.The present pedestrian detection research based on machine learning mainly concentrated in two aspects: based on statistical learning method,on the other hand is based on convolutional neural network method.Convolutional neural network has its irreplaceable advantages and is favored by the many researchers,but due to the lack of che chip solutions,the research on pedestrian detection mainly stayed in the algorithm research in Lab,or PC simulation.While the statistical learning method,through long years of research and development,has formed a series of mature theories and algorithms,the chip solution is already applied in vehicle.Therefore,the pedestrian detection based on statistical learning method is the first choice for recently pedestrian detection products in ADAS.A vehicle pedestrian detection system used in car which is based on Visconti2 7502 is proposed in this paper.Firstly,to collect many pedestrian positive and negative samples from the real road scene;then use of multi-step SVM classification to train the samples for getting dictionary file which used for determine the target in the picture is pedestrian or not.Using the camera in front of the car to capture real-time traffic information,as to different target image size and distance,we establish the image pyramid,at the same time,set ROI area for each layer of the image pyramid,and remove the useless area in the picture,in order to improve the detection speed;Finally,using two kinds of Co-occurrence Histograms of Oriented Gradient which the template size is different to extract the ROI's feature in each layer of the image pyramid.Using the dictionary file to judge the extracted target feature twice for improving the detection accuracy.Using Harris angle and optical flow method to track the pedestrian target,and realize the system by multi-core processor and image processing accelerator that integrate in the Visconti2 7502.Finally,the system is installed in the front windshield of the car at a fixed position and tested on road.Draw the conclusion: the system combins the Co HOG feature extraction algorithm with multi-step SVM classification,and implement on the Visconti2 platform.The system can achieve real-time and accurate pedestrian detection with a monocular camera.It has low cost and small volume,and is suitable for narrow space inside the vehicle.The system has great practical significance and huge market application value.