Computer Vision for Distant Vehicle Detection: How to Find Region Proposals for Low-Resolution Objects?

• Main Author: Fattal, Ann-Katrin
• Format: Others
• Language: en
• Published: 2020
• Online Access: https://tuprints.ulb.tu-darmstadt.de/11310/13/2019-10-30_Fattal_AnnKatrin.pdf; Fattal, Ann-Katrin <http://tuprints.ulb.tu-darmstadt.de/view/person/Fattal=3AAnn-Katrin=3A=3A.html> (2020): Computer Vision for Distant Vehicle Detection: How to Find Region Proposals for Low-Resolution Objects?Darmstadt, Technische Universität, DOI: 10.25534/tuprints-00011310 <https://doi.org/10.25534/tuprints-00011310>, [Ph.D. Thesis]

Safety is crucial to the development and acceptance of assisted and highly automated driving functions. In 2017, 69.3% of German fatal accidents happened on roads where the speed limit was not enforced or higher than 100km/h. At this speed, to perform safe driving maneuvers, the environment perception is a key element. Detecting objects in distances up to 200m is instrumental in anticipating potential obstacles. Due to hardware limitations, an automotive camera maps cars in e.g. 200m distance to an image of only 8px width. Hence, the absence of local details degrades the state-of-the-art detection methods designed for detecting bigger sized objects. The scope of this thesis is to develop, extend and evaluate object region localizers to improve the detection range of cameras. A saliency inspired voting map is proposed that highlights anomalies in automotive scenes. The environment is modeled with few homogeneous regions representing the background within the image. Such global features allow detecting small object regions. Inspired by the concept of learning features, this thesis presents machine learning methods detecting small objects. Existing labeled data sets such as the KITTI data set only have object regions which sizes are larger than 25px height. The presented methods in this thesis are performed against a newly created data set with 67% of object regions having a width of 8-30px, a range that has rarely been subject to research yet. Convolutional Neural Network based localizers have been evaluated and extended. To maintain a low computational power, only small networks can be used. However, such networks are limited to the usage of local features. An incorporation of global generic priors to local networks is proposed, which increases the recall especially for small object regions. The parameters to adjust Region Proposal Networks (RPNs) for the special case of small objects are further optimized and the main parameters are identified. A novel relevance based net-surgery is introduced, allowing to select the most relevant features while maintaining the recall of the RPN. It is then possible to reduce the network size to these few features.