Bounding Box Alignment Based Pedestrian Crossing Collision Avoidance Using Convolution Neural Networks

Abstract

Pedestrian detection is a challenging task for autonomous vehicles in an urban environment. Pedestrian in videos has a Variety of appearances such as occlusion and body poses and there is a proposal shift problem in pedestrian detection that cause the loss of parts such as legs and head. To address such a problem, we suggest partlevel convolution neural networks based method for pedestrian recognition using saliency map and boundary box framework in this paper. The proposed method consists of two sub-networks: person-detection and alignment. We use saliency map along with weights in the detection sub-network to remove false detections such as lamp posts and trees. The alignment network employs confidence map for better prediction of pedestrian alignment. The method is implemented and analyzed on various data sets and it hass been observed that the proposed method has better accuracy and low false positives than the existing methods.