Poster from NUS
•
0 likes•80 views
Moving object recognition (MOR) corresponds to the localization and classification of moving objects in videos. Discriminating moving objects from static objects and background in videos is an essential task for many computer vision applications. MOR has widespread applications in intelligent visual surveillance, intrusion detection, anomaly detection and monitoring, industrial sites monitoring, detection-based tracking, autonomous vehicles, etc. In this session, Murari provided a poster about the deep learning algorithms to identify both locations and corresponding categories of moving objects with a convolutional network. The challenges in developing such algorithms have been discussed.
Report
Share
Poster from NUS
- 1. WACV2018 IEEE Winter Conf. on Applications of Computer Vision Experiments MOR-UAVNet framework MOR-UAV: A Benchmark Dataset and Baselines for Moving Object Recognition in UAV Videos Murari Mandal, Lav Kush Kumar, Santosh Kumar Vipparthi Vision Intelligence Lab Malaviya National Institute of Technology Jaipur, INDIA Motivation ▪ Given a UAV video stream, how can we both localize and classify the moving objects, i.e. perform moving object recognition (MOR)? MOR-UAV Dataset ▪ About 10,948 frames are annotated with approximately 89,783 bounding boxes representing moving vehicles. These vehicles are categorized into two classes: car (80,340 bounding boxes) and heavy vehicles (9,443 bounding boxes). Table 1: Comparing MOR-UAV with other datasets Figure 3: Schematic illustration of the proposed MOR-UAVNet framework Figure 1: Difference between the three tasks: object detection (left), moving object detection (middle), and MOR (right) is depicted. Dataset Task MOR Labels Visdrone Det, T No DOTA Det No UAV123 T No UAVDT Det, T No Okutama Det, Act No Dac-sdc Det No MOR-UAV MOR Yes Figure 5: Visualization of different layers in MOR-UAVNet Table 3: Quantitative results (mAP) of the MOR-UAVNet Table 2: Inference speed, number of parameters and model size comparison of the baseline models MOR-UAV dataset attributes: - Variable object density. - Small and large object shapes. - Sporadic camera motion. - Changes in the aerial view. Future research directions: - Better mAP. - Accurately locating motion clues - Realtime challenges Method COF Vid14 Vid15 Vid16 Vid17 Overall MOR- UAVNetv1 1-3-5 82.70 40.35 53.06 17.76 48.47 1-3 86.94 32.48 85.18 29.64 58.56 1-5 53.43 32.45 91.49 06.04 45.85 1 85.65 56.41 81.35 04.98 57.09 MOR- UAVNetv2 1-3-5 71.12 19.68 69.46 29.79 47.51 1-3 79.31 40.51 59.06 31.75 52.65 1-5 83.18 38.02 80.53 30.66 58.09 1 85.57 23.34 19.17 39.02 41.77 MOR- UAVNetv3 1-3-5 39.04 35.73 16.59 49.14 35.13 1-3 60.54 25.91 05.25 17.54 27.31 1-5 71.61 33.94 61.46 08.06 43.77 1 79.04 44.09 72.85 19.27 53.81 MOR- UAVNetv4 1-3-5 60.90 32.65 79.52 14.72 46.95 1-3 65.41 48.82 38.28 29.01 45.38 1-5 80.59 41.14 62.47 19.48 50.92 1 58.47 57.62 41.67 05.71 40.87 Figure 2: Sample frames from the MOR-UAV dataset Method COF FPS #Param Model Size MOR-UAVNetv1 1-3-5 9.59 ~65.4 Million ~263.6 MB 1-3/1-5 9.59 1 11.11 MOR-UAVNetv2 1-3-5 8.44 ~36.5 Million ~146.3 MB 1-3/1-5 8.79 1 10.05 MOR-UAVNetv3 1-3-5 7.81 ~19.3 Million ~77.8 MB 1-3/1-5 8.79 1 9.59 MOR-UAVNetv4 1-3-5 9.17 ~13.2 Million ~53.3 MB 1-3/1-5 9.59 1 10.55 Figure 4: Failure cases Watch the video results here