Mihai Negru

Deep learning models are widely used for automated driving, but achieving the goal of performance in embedded Systems-on-Chip (SoCs) has a lot of challenges due to the tradeoff between accuracy and run-time. This paper addresses the task of finding a suitable neural architecture for heterogeneous Systems-on-Chip (SoCs) in order to strike a balance between accuracy and run-time efficiency. Our approach may be used to predict the inference run-time of DNN models, which is useful not only for neural architecture search, but also for the optimization of deep learning models. We propose a method for predicting the run-time of deep learning models using Graph Neural Networks, based on the Graph Convolution Network. We transform the DNN models into an acyclic graph structure, which incorporates encoded features for all nodes. Each node represents an operation of the model and all nodes are linked together. The experimental results show that our approach achieves accurate run-time inference predictions, with an overall error between 2% and 4% for ARM CPUs.

Optical flow plays an important role in many applications, such as autonomous driving, where it facilitates the perception and understanding of the surrounding environment. Since the rise of deep learning approaches for optical flow estimation, much effort has been expended on new architectures and on increasing the accuracy of predicted flow. However, topics such as including semantic knowledge in model training or estimating the uncertainty for the predicted optical flow have received comparatively little attention so far. In this paper we investigate the effect of adding semantic supervision for a state of the art supervised optical flow model [1], as well as the necessary adaptations needed for obtaining a measure of flow uncertainty. In addition, our method is suitable for deployment and integration on various embedded hardware platforms for automated driving. In our experiments, we find that semantic guidance is able to improve the model performance on the KITTI benchmark by 9% on foreground objects, after having been trained only on synthetic data from the target domain. We also obtain improved flow uncertainty estimates over commonly used approaches such as model ensembles and Monte Carlo Dropout.

A real problem of today’s society is the loss of human lives due to road accidents and the pollution caused by freight transport through metropolitan areas. The restrictions imposed in the near future for freight transport could reduce its efficiency and create many more problems. Using data centralization and developing applications or algorithms dedicated to the freight transportation sectors, routes and emissions can be managed much more efficiently. In this work, general aspects are presented, as well as a route optimization model for freight transport, taking into account the environmental impact, based on a heuristic algorithm, that of the ant colony (ACO). A multitude of studies has focused on what represents the benefits created by the applicability of solutions rather than on generalities and perspectives. The paper aims to highlight the usefulness of an optimization model of freight transport routes and the minimization of time and social costs. The study will show us that an optimized route for freight transport has a huge impact on costs, but also on time efficiency and polluting emissions.

As the technological advancement in the automotive field increases and the complexity of vehicle and infrastructure applications is extremely high, new directions and approaches are needed in this field. Supporting and developing vehicular applications dedicated to road safety by analyzing the current behavior of existing networks in various forms is imperative. This paper studies and implements a DSRC-type communications infrastructure that receives a set of controllable and adjustable indicators, which can provide messages to network drivers in a timely manner. The implementation is based on the 802.11p protocol and initially addresses pedestrian infrastructure or pedestrian safety, controlled areas, and perimeters that allow intelligent communications. The design and setting of the communication parameters in the lower layer of the DSRC stack for vehicle applications are part of this work, aspects that are also relevant in the case of autonomous vehicles.

This paper proposes two novel metrics for evaluating the consistency of object detection systems on continuous video sequences. Its aim is to help differentiate between two such systems that have similar results if evaluated with traditional metrics, so two recomputed values for precision and recall are described. They are based on a continuity score, which received its name based on the fact that it favors detections that cover more grouped frames, rather than sparse sets of detections, for true positive detections. The opposite can be said about false positives, where fragmented ones are preferred. It relies on giving bonuses to detections that appear in consecutive frames, it allows a number of interruptions in such a series, before penalties are applied, and it is designed to be highly configurable. A procedure for verifying whether two different detections represent the same object is also described. Having this correlation method, the metrics can be used for simple object detection systems, but it can also be modified such that they are used with systems that assign ids to detected objects.

Automated machine learning (AutoML) represents one of the fastest-growing fields in both industry and academia. Its popularity has increased over the last few years due to the rising demand for machine learning systems usable by non-experts. Existing open-source systems fail to fully automate the machine learning tasks, being more like a black box that generates non-reusable pipelines. We propose an end-to-end system accessible to non-specialists and data scientists who wish to effortlessly create reusable and easy to understand pipelines for any regression and classification problems. The proposed solution increases the automation level by performing an advanced data cleaning followed by a complete feature engineering process without any domain knowledge about the data set. These processes are integrated into a search that generates for each model the optimum instance of preprocessed data. Bayesian optimization (BO) with a Gaussian process (GP) maximizes the performance of these models that will be ensembled using standard and probability-based stacking. We prove that the system performs better than 43.47% of practitioners on real-world data sets, being robust to any variation of tabular data. The true potential is revealed by surpassing at least one similar tool in 50% of performed benchmarks. These results were achieved for any tested data set, regardless of its dimensionality, after a maximum of 30 minutes.

The early recognition and understanding of the actions performed by pedestrians in traffic scenes leads to an anticipation of pedestrian intentions in advance and helps in the process of collision warning and avoidance in the context of autonomous vehicles. An environment with low visibility conditions such as night-time, fog, heavy rain or smoke increases the number of difficult situations in traffic. A complete and original model for assessing if a pedestrian is engaged in a street cross action using only infrared monocular scene perception is proposed in this paper. The assessment of a street cross action is done by the time series analysis of features like: pedestrian motion, position of pedestrians with respect to the drivable area and their distance with respect to the ego-vehicle. The extraction of these features emerges from the combination of a deep learning based pedestrian detector with an original tracking algorithm, a semantic segmentation of the road surface and a time series long-short term memory network based action recognition. In order to validate the proposed method we introduce a new dataset named CROSSIR. It is formed of pedestrian annotations, action annotations and semantic labels for the road. The CROSSIR dataset is suitable for several common computer vision algorithms: (1) pedestrian detection and tracking algorithms because each pedestrian has a unique identifier over the frames in which it appears; (2) pedestrian action recognition; (3) semantic segmentation of the road pixels in the infrared image.