We introduce a novel set of features for robust object recognition. Each element of this set is a complex feature obtained by combining position- and scale-tolerant edge-detectors over neighboring ...positions and multiple orientations. Our system's architecture is motivated by a quantitative model of visual cortex. We show that our approach exhibits excellent recognition performance and outperforms several state-of-the-art systems on a variety of image datasets including many different object categories. We also demonstrate that our system is able to learn from very few examples. The performance of the approach constitutes a suggestive plausibility proof for a class of feedforward models of object recognition in cortex.
Use of bicycles on a large scale, encouraged in the context to develop an eco-friendly environment, is facing today a range of barriers. One of these barriers identified by researchers and ...governments is observed to include 'road safety'. Hence, it is necessary to set up a protection system for bicyclists especially for the cephalic segment. Currently only few studies are available concerning the head impact loading in case of real accidents. Therefore, the objective of this work is to improve the knowledge of bicyclist head trauma by following a methodology to reconstruct real-world accidents. A step is to identify the initial condition of head impact in case of real accidents. Head impact velocity and head impact area are extracted and implemented in the last generation of head injury prediction tool to simulate the head trauma by impacting directly the Strasbourg University Finite Element Head Model (SUFEHM) on the vehicle structures. The present study can be divided into three activities, i.e. obtain real bicyclist accidents' data issued from in-depth accident investigation databases, reconstruct cyclist body kinematics to obtain the initial conditions of the head just before the impact, and simulate head impact to evaluate the head loading during impact and the injury risk. A total of 24 bicyclists' accident cases with head injuries have been selected from both French and German accident databases. For each accident case, body kinematics has been simulated using Madymo® software. Two human multi-body models were used: 8 accident cases have been reconstructed by IFSTTAR using its owned developed human model and 18 accident cases have been reconstructed by Unistra using the human pedestrian TNO model. The results show that head is impacted more often on top parietal zone, and the mean impact velocity is 6.8 ± 2.7 m/s with 5.5 ± 3.0 m/s and 3.4 ± 2.1 m/s for normal and tangential components, respectively. Among these real accidents, 19 cases were selected for finite element computations by coupling the human head model and a windscreen model whose properties were extracted from literature. All reconstructed head impact gave results in accordance with the damage actually incurred to the victims.
•First application of a spatial filter to ECoG data.•The ECoG inverse problem turns out to be numerically stable.•Beamformers give reliable estimates of neural current even for dipoles far from the ...grid.•The reconstructed brain activity seems consistent with monkey and human studies.
Electrocorticography (ECoG) measures the distribution of the electrical potentials on the cortex produced by the neural currents. A full interpretation of ECoG data requires solving the ill-posed inverse problem of reconstructing the spatio-temporal distribution of the neural currents. This study addresses the ECoG source modeling developing a beamformer method.
We computed the lead-field matrix by using a novel routine provided by the OpenMEEG software. We performed an analysis of the numerical stability of the ECoG inverse problem by computing the condition number of the lead-field matrix for different configurations of the electrodes grid. We applied a Linear Constraint Minimum Variance (LCMV) beamformer to both synthetic data and a set of real measurements recorded during a rapid visual categorization task.
For all considered grids the condition number indicates that the ECoG inverse problem is mildly ill-conditioned. For realistic SNR we found a good performance of the LCMV algorithm for both localization and waveforms reconstruction.
The flow of information reconstructed by analyzing real data seems consistent with both invasive monkey electrophysiology studies and non-invasive (MEG and fMRI) human studies.
Despite a growing interest from the neuroscientific community, solving the ECoG inverse problem has not quite yet reached the level of systematicity found for EEG and MEG. Starting from an analysis of the numerical stability of the problem we considered the most widely utilized method for modeling neurophysiological data based on the beamformer method in the hope to establish benchmarks for future studies.
The enhancement of pedestrian safety to avoid traffic accidents represents a major challenge. This study allows a better understanding of the issues in pedestrian protection. It highlights the ...potential of in-depth accident studies for identifying relevant crash parameters in the pedestrian active safety. A computational simulation tool was developed to reconstruct 100 pedestrian real-world crashes. Two of them are detailed to illustrate the methodology.
A description of the complete sample is then presented which highlights the major factors affecting the detection of the pedestrian. These main factors concern the travel and impact speed of the vehicle, the pedestrian trajectory and his walking speed, the scene configuration with obstacles, and the weather conditions. In particular, it has been shown that 1 s before the impact, only 30% of pedestrians are located in front of the car and 90% of them are less than 20 m from the front of the car.
This paper highlights the potential impact points of a child pedestrian during a crash with the front end of a vehicle. Child anthropometry was defined for ages between 3 and 15 years. It was based ...on the measurement of seven different segment body heights (knee, femur, pelvis, shoulder, neck, chin, vertex) performed on about 2,000 French children. For each dimension, the 5
th, 50
th and 95
th percentile values were reported, and the corresponding linear regression lines were given. Then these heights were confronted with three different vehicle shapes, corresponding to a passenger car, a sport utility vehicle and a light truck, to identify impact points. In particular, we show that the thigh is directly hit by the bumper for children above 12 years of age, whereas the head principally impacts the hood. The influence of child anthropometry on the pedestrian trajectory and the comparison with test procedures in regulation are discussed.
With nearly one billion online videos viewed everyday, an emerging new frontier in computer vision research is recognition and search in video. While much effort has been devoted to the collection ...and annotation of large scalable static image datasets containing thousands of image categories, human action datasets lag far behind. Current action recognition databases contain on the order of ten different action categories collected under fairly controlled conditions. State-of-the-art performance on these datasets is now near ceiling and thus there is a need for the design and creation of new benchmarks. To address this issue we collected the largest action video database to-date with 51 action categories, which in total contain around 7,000 manually annotated clips extracted from a variety of sources ranging from digitized movies to YouTube. We use this database to evaluate the performance of two representative computer vision systems for action recognition and explore the robustness of these methods under various conditions such as camera motion, viewpoint, video quality and occlusion.
•Cyclist accidents cases are allocated as follows: 33% Crossing Nearside, 22% Crossing Farside, 5% Longitudinal and 34% for Turning Right and Left.•A 60 °Field Of View (FOV) (total angle of 120°) and ...a 35 m range allow detecting most cyclists.•Little gain is observed for detection rates for last time to brake higher than 1 s before collision.•51% of cyclists could be detected up to 4 s before the last time to brake with Field Of View (FOV) of 60°.
The purpose of this study was to analyze car-to-cyclist accidents to determine the challenges for an active safety system on car to avoid accidents. Based on 2261 car-to-cyclist accidents provided by in-depth accident databases, accidents are analyzed more specifically from kinematic reconstructions. The main accident scenarios are determined: crossing nearside, crossing farside, longitudinal, turning (right and left) and others. Proportion of brakes activation by the drivers before the impact was also given for those scenarios. The relative positions of the cyclists to the vehicle are analyzed from few seconds before the impact until the crash. It is observed that one second before the impact most of the cyclists were at a lateral distance smaller than 5 m to the center line of the car and less than 20 m ahead of car front. Finally, the possible detection of the cyclist by implemented sensors in the vehicle and the possible triggering of an active safety system like an Automatic Emergency Braking or a Forward Collision Warning are studied. Required detection sensors parameters, such as Field Of View (FOV) and the detection range, were analyzed relatively to the scenario’s characteristics, e.g. remaining time after cyclist appearance and before the collision, differences between scenario types. Different sensor FOVs and detection ranges were analyzed to determine their possible rates of cyclist detection. The study concluded that a FOV of 60° and a range of 35 m would detect most of the cyclists in car-to-cyclist accident scenarios. It was also concluded that in about 80% of cases, the last time to trigger brake (tLTTB), i.e. the last moment to brake in order to avoid the accident based on physical and comfort braking limitations by the car, was 1 s before the collision. It was also calculated that with a FOV of 60°, 51% of cyclists could be detected up to 4 s before tLTTB, and 72% up to 1 s before tLTTB. Values found in this paper can be useful to determine some of the specifications of an Advanced Driver Assistance System (ADAS), e.g. detection sensor coverages, available time to trigger an autonomous emergency braking or forward collision warning device. Those values are given for general ADAS sensors specification but also per scenario in the case of sensors that can adapt to a specific scenario.
The present document describes the functioning principles of the Muon Survey Tomography based on Micromegas detectors for Unreachable Sites Technology and its distinguishing features from other ...Micromegas-like detectors. Additionally, it addresses the challenges found while operating the first generation and the resulting improvements. Currently, the project Temporal Tomography of the Densitometry by the Measurement of Muons is focused on obtaining a reliable pulse from the micromesh, associated to the passing of a muon, in order to trigger the acquisition and operate in standalone mode. An outlook of the future steps of the project is provided as well.
•Kinematics of 100 real vehicle–pedestrian accidents were in-depth reconstructed.•A generic on-board active safety system with pedestrian detection was evaluated.•A system with a 35° field of view ...seems relevant for detection and crash avoidance.•A processing time to trigger an autonomous emergency braking must be less than 1s.
The purpose of this study was to analyze real crashes involving pedestrians in order to evaluate the potential effectiveness of autonomous emergency braking systems (AEB) in pedestrian protection. A sample of 100 real accident cases were reconstructed providing a comprehensive set of data describing the interaction between the vehicle, the environment and the pedestrian all along the scenario of the accident. A generic AEB system based on a camera sensor for pedestrian detection was modeled in order to identify the functionality of its different attributes in the timeline of each crash scenario. These attributes were assessed to determine their impact on pedestrian safety. The influence of the detection and the activation of the AEB system were explored by varying the field of view (FOV) of the sensor and the level of deceleration. A FOV of 35° was estimated to be required to detect and react to the majority of crash scenarios. For the reaction of a system (from hazard detection to triggering the brakes), between 0.5 and 1s appears necessary.