In this work, we propose CyDet, a hybrid approach to detection and classification based on human knowledge of the scenario context and statistical machine learning approaches to object detection, classification, and tracking. This approach enabled us to experiment with five different methods to maximize the accuracy of cyclist detection. These methods included detecting a cyclist as a single object, utilizing different classifiers for different view angles, using HOG and LBP with fewer descriptors and features, using a machine learning layer for classifier training, and uniquely, custom designed jerseys by Specialized Bicycles. By working closely with Specialized, we were able to ensure that designs were both optimized for computer vision algorithm detection and productizable, and we were able to train our algorithms to classify the known designs as cyclists which further improved our accuracy, especially when the cyclist is partially occluded. Our experiments show that our approach achieves state of the art accuracy on the KITTI [1] benchmark, and we establish a baseline detection rate with our approach on a new, open Specialized Bicycles Cyclist Detection Dataset that include cyclists wearing jerseys with known patterns.
CyDet: Improving Camera-based Cyclist Recognition Accuracy with Known Cycling Jersey Patterns
2018 IEEE Intelligent Vehicles Symposium (IV) ; 2143-2149
2018-06-01
3404289 byte
Conference paper
Electronic Resource
English
CYDET: IMPROVING CAMERA-BASED CYCLIST RECOGNITION ACCURACY WITH KNOWN CYCLING JERSEY PATTERNS
British Library Conference Proceedings | 2018
|Camera-based measurement of cyclist motion
SAGE Publications | 2019
|Child cyclist accidents and cycling proficiency training
Elsevier | 1979
|Exploring Cyclist-Vehicle Interaction – Results from a Naturalistic Cycling Study
Springer Verlag | 2021
|Analysing cyclist behaviour at cycling facility discontinuities using video data
DOAJ | 2018
|