Orientation Posture Based Vehicle Detection Using Patch Decided Vanishing Point

On-board-camera vehicle detection is a key of Advanced Driver Assistance Systems(ADAS).It can be widely used in many field such as self-guided vehicle,driver-assistance systems and automatic parking systems to solve safety problem and reduce driving difficulties.In this thesis,a vision-based vehicle detection using vanishing point is proposed,which could detect vehicles with different orientation using different models,and increases the accuracy.With the input image from the on-board camera,I first divide it into small square patches.Then there are two algorithms for grey image patches and colour image patches which help to find out useful patches for vanishing point vote process.After that,I use shadow region to generate candidates that might be vehicles.And with the vanishing point,orientation and scale of candidates are obtained.According to these,I put them into corresponding models which are trained offline.I also create a data set that included orientation and scale for vehicles and non-vehicles which is used to do experiment.From the result of the experiment,the method I proposed tend to be better than traditional method in efficiency and accuracy.