Autonomous driving technology relies heavily on sensor data for environment perception. Heterogeneous sensors such as lidar, radar, and camera have their own strengths and limitations. Therefore, relying on any single sensor would restrict the effectiveness of autonomous driving technology. However, integrating data from such heterogeneous sensors poses challenges due to differences in their representations. This article outlines a deep learning network aimed at designing modality-agnostic multi-modal fusion architecture. We study sensor data from different modalities and learn fine-grained representations using modality-specific feature encoders independently. Then, a multimodal attention-based network (MMA-Net) is proposed to fuse the data from heterogeneous modalities. The proposed MMA-Net fuses multi-modal sensor data by jointly exploiting the inter-modality and intra-modality relationships among camera, lidar, and radar sensors. The effectiveness of the proposed multi-modal fusion architecture is demonstrated using 2-D object detection metrics through extensive experiments on a dataset generated using the CARLA simulator.

Histopathological image segmentation is a laborious and time-intensive task, often requiring analysis from experienced pathologists for accurate examinations. To reduce this burden, supervised machine-learning approaches have been adopted using large-scale annotated datasets for histopathological image analysis. However, in several scenarios, the availability of large-scale annotated data is a bottleneck while training such models. Self-supervised learning (SSL) is an alternative paradigm that provides some respite by constructing models utilizing only the unannotated data which is often abundant. The basic idea of SSL is to train a network to perform one or many pseudo or pretext tasks on unannotated data and use it subsequently as the basis for a variety of downstream tasks. It is seen that the success of SSL depends critically on the considered pretext task. While there have been many efforts in designing pretext tasks for classification problems, there have not been many attempts on SSL for histopathological image segmentation. Motivated by this, we propose an SSL approach for segmenting histopathological images via generative diffusion models. Our method is based on the observation that diffusion models effectively solve an image-to-image translation task akin to a segmentation task. Hence, we propose generative diffusion as the pretext task for histopathological image segmentation. We also utilize a multi-loss function-based fine-tuning for the downstream task. We validate our method using several metrics on two publicly available datasets along with a newly proposed head and neck (HN) cancer dataset containing Hematoxylin and Eosin (H&E) stained images along with annotations.

We present the delta-gesture biometrics quantification assessment (DGBQA) framework which estimates the biometric characteristics of hand gestures. The proposed framework is aimed at learning generic motion-representations of gestures instead of subject-specific details from a large number of identities. It also enables the biometric scores to be estimated for a set of gestures at a time instead of having to estimate these one at a time. In the first step, it formulates a feature space which is identity and gesture aware, and in the second step, it proceeds to compute biometric scores using inter-subject and intra-subject distance measures in the feature space. However, due to the inclusion of identity-aware objective, the identity details tend to be shared across gestures. We refer to this as identity sharing and this can lead to the score for different gestures being dependent on each other. To address this issue, we introduce an identity-level cross-gesture disentanglement loss$(\mathscr{L}_{ICGD})$which encourages the different gestures belonging to the same identity to be orthogonal in the feature space. We demonstrate the efficacy of the proposed biometric quantification framework and the disentanglement loss function through extensive experiments on four datasets and using standard as well as proposed novel evaluation metrics. Our analysis indicates that gestures involving multiple coarse movements are better for biometrics.

As frequency modulated continuous wave (FMCW) radars are becoming a key component of autonomous driving technology, mutual interference among automotive FMCW radars is a significant challenge. In this work, we propose a novel signal processing solution to effectively mitigate mutual interference among FMCW radars. The proposed framework distinguishes the target echoes and interference using their characteristics in the time-frequency spectrum, both of which appear as lines but with different slope characteristics. An interference map is generated from the interference contaminated time-frequency spectrum by employing the Hough transform along with slope filtering. The interference map is then used as a mask for interference mitigation. The interference-mitigated signal is reconstructed by applying time-frequency ridge-based mapping to subjugate the power loss associated with the target signals. We carry out extensive simulations to demonstrate the interference mitigation capabilities of the proposed Hough trans-form and time-frequency ridge (HTFR)-based scheme in terms of the signal-to-interference plus noise ratio (SINR), correlation coefficient, and computational efficiency.

Received signal strength (RSS)-based localization has been popular as it can achieve reasonable accuracy without the need for additional hardware. In this work, we consider RSS-based cooperative and non-cooperative localization when the transmit power of the nodes is unknown. The maximum likelihood (ML) formulation of the RSS-based localization is non-linear, non-convex, and discontinuous and cannot be solved using conventional techniques. Firstly, we linearize the relation between pairwise distance between nodes and received power using a least squares-based linearization technique. Next, we propose two RSS-based localization techniques, referred to as SDP-URSS and SDP-RSS, that convert the ML into a constrained convex optimization problem using the proposed linearization technique and semidefinite relaxation. SDP-URSS assumes the transmit powers are unknown and estimates them along with the location of the nodes, whereas SDP-RSS uses the transmit power information to improve the location estimates. Extensive performance evaluation of the proposed technique considering various performance metrics under non-cooperative and cooperative scenarios demonstrates the superiority of the proposed localization techniques over the existing methods.

Received signal strength (RSS)-based localization techniques have attracted a lot of interest as they are easy to implement and do not require any localization-specific hardware. However, maximum likelihood (ML) formulation of RSS-based localization problem is non-convex, non-linear, and discontinuous, and cannot be solved using standard optimization techniques. We propose techniques that converts the ML objective function into an invex (invariant convex) function and solve them using gradient descent. We also employ coordinate descent to solve the invex problem in a completely distributed manner without any synchronization requirements. The coordinate descent-based technique can be implemented on the sensor nodes as it has low computational complexity and scales very well to large networks. We prove the convergence theoretically, derive the convergence rate, and provide a detailed computational complexity and communication overhead analysis of the techniques. We perform extensive performance analysis and compare our techniques with centralized and distributed localization methods, and demonstrate the superior performance of the proposed techniques in terms of convergence rate, localization accuracy, and execution time.

In many pattern recognition and computer vision applications, the observed high dimensional data often contains redundant information which can lead to increased computational complexity and overfitting. Most classification approaches learn discriminative projections with respect to the strict binary label matrix. This results in overfitting and loss of intrinsic structure of the observed data. To overcome these limitations, we propose low-rank double relaxed regression (LR-DRR) for image classification. LR-DRR optimizes the projections using a relaxed target matrix for regression based on the class label information. The proposed work extracts the projections while jointly considering a relaxed target matrix and a low-rank error term in the regression framework so as to allow double relaxation. Numerical experiments on several public data sets for face recognition, object classification, and scene classification applications show that the proposed approach is able to identify low-dimensional discriminative features and outperforms the state of the art classification methods.

Geotagging of sensor data in a wireless sensor network is important in many applications. It may not always be possible to record locations of the sensor nodes during deployment. Localization techniques provide the node location information, however most of these techniques rely on the neighboring nodes for localization. Thus it is essential that the information from neighbors be trustworthy and/or the localization techniques be robust to some of the nodes turning malicious. In this paper, we address the scenario where the malicious node(s) attempt to disrupt the localization process of a target node in an uncoordinated manner. We propose a secure localization technique, called the weighted least square (WLS) localization, in a network with one or more compromised anchor nodes (nodes with known positions are referred to as anchor nodes). The WLS technique assigns larger weights to the anchor nodes that are closer to the target node and is shown to offer significant advantages over existing techniques. The Cramer-Rao lower bound (CRLB) on the root mean square error (RMSE) of the position estimate for the uncoordinated attack is also derived. The proposed technique is shown to provide better localization accuracy than existing algorithms.

Economical, low-power, and easy to deploy indoor localization schemes have always been a challenge. We propose a technique that utilizes the analog output of pyroelectric (or passive) infrared (PIR) sensors for indoor localization. We propose two models for distance estimation based on the characterization of the analog sensor output in terms of the peak to peak value and its relationship with distance from the sensor. The proposed distance estimation models are based on bijective functions: Mbhusing a hyperbolic function and Mplusing a piecewise linear function. Once distances of the subject from the PIR sensors (or anchors) are estimated, multilateration or support vector regression (SVR) based techniques are used for computing the location coordinates of the subject. Using four PIR sensors, we demonstrate the localization of a human subject in a 7 m × 7.5 m area with the regression based technique outperforming the other techniques in terms of accuracy and achieving an RMS localization error of 0.65 m using the Mplmodel for distance estimation. We also compare the computational complexity and memory or storage requirements of the proposed techniques which are an important consideration for distributed implementation on resource constrained devices such as sensor nodes.

In this paper we present occupancy estimation techniques using real (motion, door closure) and virtual (PC activity detector) sensors. The techniques considered here are based on the decision tree and artificial neural network models. Results from an experimental test-bed in a four person office room are also presented.

Many of the physical systems in the real world together with their surrounding environments and the networks used for monitoring these systems can be mathematically modeled as a graph. We present a graph-based optimization technique for source localization in such systems. The proposed technique uses the detection times of any particular event or phenomenon of interest at different measurement points within the system to determine the actual location or source of the event. The graphical model represents the propagation characteristics of the physical phenomenon as well as the topology of the monitoring network. The source localization algorithm is validated using experimental data from a wireless sensor network test bed used for monitoring the drinking water distribution system. We also present a method for analyzing the sensitivity of the localization result to estimation errors in the parameters of the graphical model.

Sensor networks used for event detection scenarios are considered where the goal is to detect any abnormal conditions with minimum possible delay. We propose a greedy heuristic based sensor selection and sequential detection procedure that significantly improves the detection speed, measured in terms of the number of measurements needed for detection, and is suitable for distributed implementation. In the proposed model, the fusion center selects one sensor at a time for measurement while maximizing a greedy heuristic. Instead of collecting a fixed number of measurements, the fusion center collects one measurement at each time step, until by a sequential decision rule the collection stops and a decision is made. The sequential procedure significantly outperforms a non-sequential (or fixed sample size) detector in that it always needs fewer measurements on average to achieve the same detection performance.

In this paper, we present an overview of a system that we are developing for the continuous monitoring of water distribution systems in Singapore. We discuss the design and implementation of the system that enables real-time detection of pipe burst events.

We present the implementation of a directional beacon-based positioning algorithm using radio frequency signals. This algorithm allows each mobile node to compute its position with respect to a set of reference nodes which are equipped with a rotating directional antenna. The use of directional beacon-based algorithm for position location eliminates the need for strict synchronization between the reference nodes and the mobile node. In contrast to positioning algorithms that rely on signal propagation time and bandwidth, the proposed algorithm depends on the beam-width and rotational speed of the directional antenna. We will show that these parameters can be optimized with a low cost solution that provides good positioning accuracy. The system implementation is based on the GNU radio software platform and the universal software radio peripheral as the hardware component. We present an enhanced maximum likelihood method for estimating the received signal amplitude profile. To deal with obstructed line-of-sight scenarios, we do not rely purely on the received signal strength and instead formulate a least squares problem to estimate the line-of-sight component in a multipath environment. These advanced signal processing techniques yield a more accurate estimate of the bearing of the mobile node with respect to each of the reference nodes. We also show that the proposed positioning algorithm is tolerant to errors in timing and synchronization. We demonstrate the ability to obtain mobile node position estimates with sub-meter accuracy by transmitting a narrowband signal of 1 kHz bandwidth in the 2.4-2.5 GHz band. The experimental results show a mean position error of 0.759 m, in a field measuring 55m by 43m, using eight 90° rotations of the antenna.

The goal of the sensor network localization problem is to determine positions of all sensor nodes in a network given certain pairwise noisy distance measurements and some anchor node positions. This paper describes a distributed localization algorithm based on second-order cone programming relaxation. We show that the sensor nodes can estimate their positions based on local information. Unlike previous approaches, we also consider the effect of inaccurate anchor positions. In the presence of anchor position errors, the localization is performed in three steps. First, the sensor nodes estimate their positions using information from their neighbors. In the second step, the anchors refine their positions using relative distance information exchanged with their neighbors and finally, the sensors refine their position estimates. We demonstrate the convergence of the algorithm numerically. Simulation study, for both uniform and irregular network topologies, illustrates the robustness of the algorithm to anchor position and distance estimation errors, and the performance gains achievable in terms of localization accuracy, problem size reduction and computational efficiency.

The goal of the sensor network localization problem is to determine positions of all the sensor nodes in a network given certain pairwise noisy distance measurements and inaccurate anchor node positions. A two-step distributed localization approach based on second-order cone programming (SOCP) relaxation is presented. In the first step, the sensor nodes determine their positions based on local information and in the second step, the anchor nodes refine their positions using information from the neighboring nodes. Our numerical study shows that the sensor and anchor positions cannot be estimated in a single step; the sensors must be estimated first for the results to converge. The second step enables anchors which are in the convex hull of their neighbors to refine their positions. Extensive simulation results with inaccurate anchor positions and noisy distance measurements are presented. These illustrate the robustness of the algorithm and the performance gains achievable in terms of problem size reduction, computational efficiency and localization accuracy.

This paper addresses ranging in indoor quasi-static sensor environments and presents a time-of-arrival (TOA) based ranging algorithm. A statistical model of the multipath channel in the form of the signal return and noise characterization is derived, and utilized to distinguish signal components from noise. The algorithm then uses multiple signal receptions at each base station, to differentiate between line-of-sight (LOS) and non-LOS components, and to accurately estimate the position of the LOS component in the received multipath signal. The location is estimated through a mathematical programming problem formulation. Using a synthesized bandwidth of 2 GHz, a 4-bit analog-to-digital converter (ADC) and with 5-10 dB signal-to-noise ratio (SNR), range estimation with sub-meter accuracy is achieved. Furthermore, the associated range estimation error does not increase with increase in the transmitter-receiver range