Search

International audience; Conventional tracking filter in marine radar only concerns the position or shape of the object. When the target encounters the clutter interruption or ship occlusion, the tracker is easy to drift away due to the wrong target association. Since in short-range maritime surveillance, the X-band marine radar captures the object in an extended region with varying intensities, in this paper, combining position, shape, and appearance of the target together, multiple kernelized correlation filters (MKCF) are proposed to conduct a single object tracking in the real marine radar. By automatic initializing KCF on a target in different time steps and fusing these multiple KCFs via the maximum likelihood, the proposed tracker implicitly uses multiple instances of the target to improve the robustness of the long-term tracking. Bounding rectangles of the multiple trackers also enhance the reliability of ship segmentation via a voting procedure. In the real ship tracking experiment, the proposed tracker performs favorably against the conventional radar tracker and top-ranked visual tracker in the cases of clutter interruption, ship occlusions, and scale changing. We believe that our proposal sheds light on that the intensity distribution of the extended object could be valuable for target association in marine radar data. To encourage further research, our tracking framework is made open-source.

International audience; Tracking-by-detection is a popular framework for Multiple Object Tracking (MOT) where detectors produce a set of labeled detection to indicate the categories, the size, and the position of the objects. Most of the MOT approaches use the detection as points to do association. However, in this manner, these methods ignore the majority of useful association cues. As a result, the accuracy of the association is not optimal. In our approach, we combine multiple cues like object appearance feature, object size and position to improve the association confidence. Instead of treating the detection and tracking as two separate parts we directly extract the appearance features from the detector's feature map and append an extra association network to fuse the multiple cues. The architecture of our approach is an end-to-end detection-association network. The input of our network is the image sequence, and the output is the inter-frame target association matrix. We tested our network on MOT17 challenge dataset. The results show that our solution significantly improves the short term association accuracy when compared with the single-cues association methods while keeping a lower consumption of computing resources.

International audience; Abnormal event detection is a complex task in computer vision. It is one among the most difficult challenges in crowded scene analysis, given the huge variety of events to analyze and the frequent occlusions that can occur. Ideally, to meet this challenge, it would be necessary to have a large number of normal and abnormal training examples. However, the high variability between events and their complexity make it almost impossible to provide enough abnormal training samples. To overcome these constraints, the solution may reside in a robust one-class learning algorithm. In this paper, we propose a novel one-class neural network technique suitable for outliers detection. This network takes advantage of the deep learning ability in image processing, while meeting the constraint of using only training samples from the target class. Our network is composed of a Convolutional Auto Encoder (CAE) with two loss functions. The first one is a standard mean squared error (MSE) and the second is a proposed compactness loss function particularly adapted for one class learning. This network allows us to not only extract robust spatiotemporal features but also ensures that all the target class representations are sufficiently compact to be effectively isolated from outliers. The proposed algorithm is evaluated on the challenging USCD Ped2 dataset for abnormal video event detection and localization. Both qualitative and quantitative experiments show competitive performances compared to the state-of-the-art methods.