Upper Body Pedestrian Detection
Author
Thomsen, Simon Mark
Term
4. term
Publication year
2016
Submitted on
2016-06-02
Pages
55
Abstract
Trafikulykker med fodgængere indebærer en høj risiko for dødsfald. Derfor er der brug for systemer, der kan opdage fodgængere i trafikale situationer. Et almindeligt problem er tildækning (occlusion): Den nederste del af kroppen er ofte skjult, og derfor fokuserer dette arbejde på at registrere overkroppen. Den foreslåede detektor bruger konvolutionsneuronale netværk (CNN) – en type maskinlæring til billedanalyse – i en kaskade med tre trin. Det første netværk er meget lavt (få lag), det næste er moderat dybt, og det sidste er dybere og foretager den endelige vurdering. Denne trinvise opbygning skal holde beregningsbehovet nede ved hurtigt at frasortere kandidater, før der bruges mere processering på de mest lovende, så systemet bliver lettere end ét enkelt, meget dybt CNN og i princippet mere egnet til brug i virkeligheden. Forsøgene viser, at overkropsdetektoren præsterer dårligere end fuldkropsdetektorer til fodgængere. Et fald var delvist forventeligt, fordi en overkrop rummer mindre information end en hel person, men forskellen er stor nok til, at systemet endnu ikke er anvendeligt i praksis. For at gøre det brugbart kræves yderligere forskning i bedre udtrækning af kendetegn fra overkroppe.
Traffic accidents involving pedestrians carry a high risk of fatality. This creates a need for systems that can detect pedestrians in traffic scenes. A common challenge is occlusion: the lower part of the body is often hidden, so this work focuses on detecting the upper body. The proposed detector uses convolutional neural networks (CNNs)—a type of machine learning for image analysis—in a three-stage cascade. The first network is very shallow (few layers), the second is moderately deep, and the last is deeper and makes the final prediction. This staged design aims to keep computation low by quickly filtering candidates before spending more processing on the most promising ones, making the system lighter than a single very deep CNN and, in principle, more suitable for real-world use. Experiments show that the upper-body detector performs worse than full-body pedestrian detectors. While some drop was expected because an upper body contains less information than a whole person, the gap is large enough that the system is not currently usable in practice. Further research is needed to improve how upper-body features are extracted.
[This abstract was generated with the help of AI]
Keywords
Documents