Unsafe pose recognition using convolutional spatial pyramid and boosted random forests

Abstract

Pedestrian pose recognition is important work for early accident prevention in advance driver assistance system (ADAS). In particular, because most pedestrian-vehicle crashes are occurred from late of night to early of dawn, our study focus on recognizing unsafe pose using thermal image captured from moving vehicle. For feature extraction from images, we apply histogram of gradient (HOG) and oriented center symmetric local binary patterns (OCS-LBP) to the input image as convolutional filters. Then the feature maps are generated from convolutional filters and we apply spatial pyramid pooling to the feature maps to extract global and local feature together and maintain spatial information by pooling local spatial bins. After feature extraction, all feature descriptors are aggregated as one descriptor and it is applied to boosted random forest to classify pedestrian poses with small number of decision trees. Boosted random forest maintain generality with small number of decision trees by using the fact that sequential training constructs complementary decision trees for the training samples. To make the unsafe poses that can encounter while driving a vehicle, we define six poses such as ‘standing’, ‘running’, ‘walking’, ‘sitting’, ‘collapsing’, and ‘view-back’ and collect 150 test images containing individual pose. The proposed algorithm is successfully applied to test thermal images and showed good performance on six different poses.