PEDESTRIAN DETECTION USING MOVING CAMERA

Abstract

A perfect on-board Pedestrian Detection System, referred to as PDS, must detect the presence of people, stationary or moving, on the way of the vehicle and react according to the risk of running over the pedestrian. The action performed by PDS in case a pedestrian appears right before the moving vehicle and is likely to get harmed includes: warning the driver in advance, or apply braking action, or deploy external airbags, perform an evasive manoeuvre or else. It is also necessary that the entire system works well without disturbing the driver needlessly in normal situations, if there is no risk at all. In addition to that, such a system should work satisfactorily well independent of the time, road, and weather conditions. Additionally, the cost of the pedestrian detection module should be relatively small compared to the total cost of the vehicle. Following are the various modules and techniques based on computer vision used that together constitute a Pedestrian Detection System, added to that various methods used to make it suitable for our application involving pedestrian detection using moving camera. For Pedestrian Segmentation, sliding window based approach will be used. As the camera is moving, then in such moving systems case the prevalent methods for segmentation like background subtraction will get failed. Therefore, with the help of sliding window technique which scans through the frame through different scales, segmentation of the object of interest is done. For feature extraction, HOG (Histogram of Oriented Gradients) based features descriptor is favourable because it is independent of intensity and based on change in intensity levels i.e. gradients or edges. Moreover, it works pretty well when it comes to dealing with human detection related applications in particular. After feature extraction, a suitable Neural network classifier modelled on training and subsequent testing, decides whether the region cropped by sliding window is in actual containing a pedestrian or not. So here basically the classifier is trained for four classes. These four classes correspond to four different view of pedestrian like: front view, side view (left or right) and rear view. For training, features of positive and negative images are extracted. Positive images are pedestrian images and negative images are images of possible background objects. In addition to that, Neural network is preferred over SVM classifier because single neural classifier can be used for training and classification of multiple classes and for large set of database convergence is better in neural network based classifier.