#304: Multi-cue vehicle detection for semantic video compression in georegistered aerial videos

N. Al-Shakarji, F. Bunyak, H. AliAkbarpour, G. Seetharaman, and K. Palaniappan

Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pgs. 56-65, 2019

deep learning, object detection, data compression, video surveillance, image motion analysis

PlainText, Bibtex, PDF, URL, Google Scholar


Detection of moving objects especially vehicles in videos acquired from an airborne camera is very useful for video analytics applications including traffic flow, urban planning, surveillance, law enforcement and disaster response. Using fast low power algorithms for onboard moving object detection would also provide region of interest-based semantic information for very high image compression. Despite recent advances in both UAV platforms and imaging sensor technologies, vehicle detection from aerial video remains challenging due to the relatively small object sizes, appearance changes, platform motion and camera jitter, obscurations and the scene and environment complexity. This paper proposes an approach for moving vehicle detection which synergistically fuses both appearance and motionbased detections in a complementary manner using deep learning combined with flux tensor spatio-temporal filtering [28]). We use deep learning as an appearance-based approach to detect basically all vehicles (both moving and stationary) present in the scene. For detecting moving objects a spatiotemporal filtering is used (Flux tensor [28]) which detects any type of motion including real moving objects and also spurious motions (i.e. parallax motions caused by buildings and non-flat scene structures and magnified by the platform motion). Our proposed pipeline is able to detect the moving vehicles and filter out the false positives caused by parked cars, through fusion of both appearance and motion based techniques. Experimental results show the effectiveness of the proposed method.