MACHINE PERCEPTION AND ROBOTICS GROUP

機械知覚&ロボティクスグループ
中部大学

Pedestrian detection is a necessary technology to ensure safety in automated driving. In order to realize a safer and more secure automated driving, it is necessary to clarify the basis of the pedestrian detection. Conventional pedestrian detection methods use Convolutional Neural Network (CNN) object detection to detect pedestrians. Several methods use Region Proposal Network (RPN), which predicts the position of a pedestrian by determining the offset between the pedestrian and a set of rectangular boxes called anchors. RPNs compute coordinates and size offsets, there are other areas to focus on besides pedestrians. However, they are not clear as a basis of judgment for pedestrian detection. In this study, we visualize the basis of judgment for pedestrians by using Pixel-wise-Attention in the Point-base object detection method, which detects without anchors. This method introduces a Pixel-wise-Attention mechanism to obtain the attention map for an arbitrary pixel in the Center and Scale Prediction (CSP), which detects pedestrians based on the center of the detection target and the scale from that point. By acquiring the attention map of the center of the estimated pedestrian, we can obtain pedestrians and the surrounding area of interest as a basis for judgment. In our experiments, we visualize the attention area of the CSP with the Pixel-wise-Attention mechanism and show the basis for the decision of pedestrian detection.