Abstract: In order to solve the problem of low detection accuracy of pedestrians with small stature or in complex background in advanced driver assistance systems (ADAS), a real-time pedestrian detection model named YOLOv2-P for ADAS is established based on a deep neural network model YOLOv2. Firstly, the parametric rectified linear unit activation function in feature extraction network is adopted to adaptively learn parameters from training data, and the multi-feature map fusion method is used to fusion low-layer feature map and high-layer feature map in feature extraction network. Next, cross entropy loss function is used instead of sigmoid activation function in model YOLOv2, and the width and height loss functions are normalized. Finally, a clustering is performed on the pedestrian boundary frame size in pedestrian data set by utilizing iterative self-organizing data analysis algorithm. The results of test show that the detection accuracy of the pedestrians with small stature or in complex background with YOLOv2-P model has a significant rise, compare with that with YOLOv2 model, meeting the requirements in accuracy and real-time performance for the pedestrian detection in ADAS.

Key words: pedestrian detection, advanced driver assistance system, parametric rectified linear unit, cross entropy loss function, iterative self-organizing data analysis algorithm