• We designed a novel architecture for video object detection that capitalizes on temporal information. fusion module to merge feature maps coming from several temporally close frames. proposed an improvement the SpotNet attention module. trained and evaluated our with three different base detectors two traffic surveillance datasets. demonstrated consistent significant of model over baselines. ...

In this paper, we propose a novel framework for behaviors recognition and detection of certain types of abnormal behaviors, capable of achieving high detection rates on a variety of real-life scenes. The new proposed approach here is a combination of the location based methods and the object based ones. First, a novel approach is formulated to use optical flow and binary motion video as the loc...

Most popular graph attention networks treat pixels of a feature map as individual nodes, which makes the embedding extracted by convolution lack integrity object. Moreover, matching between template and search using only part-level information usually causes tracking errors, especially in occlusion similarity situations. To address these problems, we propose novel end-to-end framework that has ...

Multispectral image pairs can provide the combined information, making object detection applications more reliable and robust in open world. To fully exploit different modalities, we present a simple yet effective cross-modality feature fusion approach, named Cross-Modality Fusion Transformer (CFT) this paper. Unlike prior CNNs-based works, guided by transformer scheme, our network learns long-...

3D object detection with LiDAR and camera fusion has always been a challenge for autonomous driving. This work proposes deep neural network (namely FuDNN) LiDAR–camera detection. Firstly, 2D backbone is designed to extract features from images. Secondly, an attention-based sub-network fuse the extracted by point clouds PointNet++. Besides, FuDNN, which uses RPN refinement of PointRCNN obtain bo...

In object detection, non-maximum suppression (NMS) methods are extensively adopted to remove horizontal duplicates of detected dense boxes for generating final instances. However, due the degraded quality detection and not explicit exploration context information, existing NMS via simple intersection-over-union (IoU) metrics tend underperform on multi-oriented long-size objects detection. Disti...