Search

We present a new simulation framework for learning control policies for autonomous vehicles (AVs) based on real-world vehicle data and maps. The framework we propose consists of three major components: a) creating a detailed lane-level map (i.e., a high definition (HD) map) for a region of interest, b) generating an environment on the HD-map using the sensor outputs (e.g., GPS, radar) from vehicles driving in the same region, and c) learning a control policy based on the realistic environment constructed. We created the lane-level HD-maps using open street maps (OSM) and aerial imagery, from which we extracted the lane-level marking and edge features. The extracted image features are then utilized to calculate higher level attributes (e.g., curvature, heading, cross-sections etc.) for each point in the HD-map. The data acquired from vehicle sensors is combined with the constructed map to create a realistic environment. Based on the constructed environment, we learned a policy to control the vehicle laterally using a reinforcement learning algorithm and longitudinally using proportional-integral-derivative (PID) controller. Our experimental results show that the proposed framework works well, offering a flexible and scalable solution for learning control policies for AVs in realistic environments.

By exchanging basic traffic messages among vehicles for safety applications, a significantly higher level of safety can be achieved when vehicles and designated infrastructure-locations share their sensor data. While cameras installed in one vehicle can provide visual information for mitigating many avoidable accidents, a new safety paradigm is envisioned where visual data captured by multiple vehicles are shared and fused for significantly more optimized active safety and driver assistance systems. The sharing of visual data is motivated by the fact that some critical visual views captured by one vehicle or by an infrastructure-location are not visible or captured by other vehicles in the same environment. Sharing such data in real-time provides an invaluable new level of awareness that can significantly enhance a driver-assistance, connected vehicle, and/or autonomous vehicle's safety-system.

This paper addresses the problem of identifying a very small subset of data points that belong to a significantly larger massive dataset (i.e., Big Data). The small number of selected data points must adequately represent and faithfully characterize the massive Big Data. Such identification process is known as representative selection [19]. We propose a novel representative selection framework by generating an l 1 norm sparse graph for a given Big-Data dataset The Big Data is partitioned recursively into clusters using a spectral clustering algorithm on the generated sparse graph. We consider each cluster as one point in a Grassmann manifold, and measure the geodesic distance among these points. The distances are further analyzed using a min-max algorithm [1] to extract an optimal subset of clusters. Finally, by considering a sparse subgraph of each selected cluster, we detect a representative using principal component centrality [11]. We refer to the proposed representative selection framework as a Sparse Graph and Grassmann Manifold (SGGM) based approach. To validate the proposed SGGM framework, we apply it onto the problem of video summarization where only few video frames, known as key frames, are selected among a much longer video sequence. A comparison of the results obtained by the proposed algorithm with the ground truth, which is agreed by multiple human judges, and with some state-of-the-art methods clearly indicates the viability of the SGGM framework.

Partial Differential Equation (PDE) based diffusion has been utilized for image denoising for more than two decades. It is known that the process of diffusion preserves the edges and object boundaries making it a suitable preprocessing step for edge detection. Synergetic to these efforts, in this work, we apply diffusion to network graphs leading to an efficient algorithm for removing anomalous structures from the network topology. The driving force for the diffusion process in our work is the clustering propensity that exists in real social networks. The proposed diffusion enhances the boundaries of communities of which makes it a suitable step prior to community detection.

In this paper we consider the problem of recovering an N-dimensional data from a subset of its observed entries. We provide a generalization for the smooth Shcatten-p rank approximation function in [1] to the N-dimensional space. In addition, we derive an optimization algorithm using the Augmented Lagrangian Multiplier in the N-dimensional space to solve the tensor completion problem. We compare the performance of our algorithm to state-of-the-art tensor completion algorithms using different color images and video sequences. Our experimental results showed that the proposed algorithm converges faster (approximately half the execution time), and at the same time it achieves comparable performance to state-of-the-art tensor completion algorithms.