No abstract

In this paper, we present methods for segmenting noisy two-dimensional forward-scan sonar images and classify and model their background. The segmentation approach differentiates the highlight blobs, cast shadows, and the background of sonar images. There is usually little information within relatively large background regions corresponding to the flat sea bottom and (or) water column, as they are often corrupted with speckle noise. Our experiments show that the background texture is dominated by the speckle noise which has the appearance of a pseudo-random texture. We show that the background texture of the underwater sonar images can be categorized by a small number of classes. The statistical features work better than the texture-based features in categorizing the pseudo-random background, which further strengthen our hypothesis of the dominance of noise over the background texture. As a result, we can model the noisy background with a few parameters. This has an application in coding the sonar images in which highlight blob regions and cast shadows are coded at the encoder side while the speckle noise-corrupted background can be synthesized at the decoder side. Since the background regions occupy a large fraction of the FS sonar image, we expect higher compression rates than most current image or video coding standards and other custom-designed sonar image compression techniques.Index Terms-forward-scan sonar imagery, sonar image segmentation, sonar background classification, speckle noise modeling and synthesis.

This paper addresses the problem of computational efficiency with respect to Dynamic Programming (DP) based Track-Before-Detect (TBD). Generally, DP-TBD is a grid-based method that estimates target trajectory by means of searching all the physically admissible paths in a finite discrete state space. However, its computational complexity increases nonlinearly as the state space expands with size or dimension, which heavily restricts its applications for many realistic scenarios, e.g., radar target detection. To alleviate this problem, a fast twostep implementation of DP-TBD is proposed in this work. In the first step, to reduce the computational costs, target states are roughly estimated by searching a discrete grid with larger cell size. In the second step, an accurate search is performed only on the parts of the state space indicated by the results in the first step with reasonable computational cost. Additionally, the proposed algorithm is also suitable for the surveillance radar scenarios. Simulation results show that the proposed algorithm can effectively improve the computational efficiency of DP-TBD with limited performance degradation.

No abstract

Multi-modality fusion is the guarantee of the stability of autonomous driving systems. In this paper, we propose a general multi-modality cascaded fusion framework, exploiting the advantages of decision-level and feature-level fusion, utilizing target position, size, velocity, appearance and confidence to achieve accurate fusion results. In the fusion process, dynamic coordinate alignment(DCA) is conducted to reduce the error between sensors from different modalities. In addition, the calculation of affinity matrix is the core module of sensor fusion, we propose an affinity loss that improves the performance of deep affinity network(DAN). Last, the proposed step-by-step cascaded fusion framework is more interpretable and flexible compared to the end-toend fusion methods. Extensive experiments on Nuscenes [2] dataset show that our approach achieves the state-of-theart performance.dataset show that our approach achieves the state-of-the-art performance.