Vehicle detection in wide area motion imagery (WAMI) has drawn increasing attention from the computer vision research community in recent decades. In this paper, we present a new architecture for vehicle detection on road using multi-task network, which is able to detect and segment vehicles, estimate their pose, and meanwhile yield road isolation for a given region. The multi-task network consists of three components: 1) vehicle detection, 2) vehicle and road segmentation, and 3) detection screening. Segmentation and detection components share the same backbone network and are trained jointly in an end-to-end way. Unlike background subtraction or frame differencing based methods, the proposed Multitask Assessment of Roads and Vehicles Network (MARVN) method can detect vehicles which are slowing down, stopped, and/or partially occluded in a single image. In addition, the method can eliminate the detections which are located at outside road using yielded road segmentation so as to decrease the false positive rate. As few WAMI datasets have road mask and vehicles bounding box anotations, we extract 512 frames from WPAFB 2009 dataset and carefully refine the original annotations. The resulting dataset is thus named as WAMI512. We extensively compare the proposed method with state-of-the-art methods on WAMI512 dataset, and demonstrate superior performance in terms of efficiency and accuracy.
Multitask assessment of roads and vehicles network (MARVN)
Sensors and Systems for Space Applications XI ; 2018 ; Orlando,Florida,United States
Proc. SPIE ; 10641
2018-05-02
Aufsatz (Konferenz)
Elektronische Ressource
Englisch
Multitask assessment of roads and vehicles network (MARVN) [10641-12]
British Library Conference Proceedings | 2018
Engineering Index Backfile | 1910
|Safer roads with automated vehicles?
TIBKAT | 2018
|Heavy vehicles and light roads
Engineering Index Backfile | 1933
|Semi-autonomous Vehicles for Indian Roads
Springer Verlag | 2024
|