Automatic Multi-Scale and Multi-Object Pedestrian and Car Detection in Digital Images Based on the Discriminative Generalized Hough Transform and Deep Convolutional Neural Networks

Many approaches have been suggested for automatic pedestrian and car detection to cope with the large variability regarding object size, occlusion, background variability, aspect and so forth. Current state-of-the-art deep learning-based frameworks rely either on a proposal generation mechanism (e.g., "Faster R-CNN") or on the inspection of image quadrants / octants (e.g., "YOLO" or "SSD"), which are then further processed with deep convolutional neural networks (CNN). In this thesis, the Discriminative Generalized Hough Transform (DGHT), which operates on edge images, is analyzed for the application to automatic multi-scale and multi-object pedestrian and car detection in 2D digital images. The analysis motivates to use the DGHT as an efficient proposal generation mechanism, followed by a proposal (bounding box) refinement and proposal acceptance or rejection based on a deep CNN. The impact of the different components of the resulting DGHT object detection pipeline as well as the amount of DGHT training data on the detection performance are analyzed in detail. Due to the low false negative rate and the low number of candidates of the DGHT as well as the high classification accuracy of the CNN, competitive performance to the state-of-the-art in pedestrian and car detection is obtained on the IAIR database with much less generated proposals than other proposal-generating algorithms, being outperformed only by YOLOv2 fine-tuned to IAIR cars. By evaluations on further databases (without retraining or adaptation) the generalization capability of the DGHT object detection pipeline is shown.