Effective traffic monitoring systems must take into account all moving objects. While vehicle detection systems are common, sensor-based solutions that provide awareness of pedestrian activity have yet to become an integral part of a smart and effective infrastructure capable of protecting these most vulnerable traffic participants. Such a system could alert incoming vehicles about dangerous situations involving pedestrians, or provide adaptive traffic light control systems with information about the motion of people, so that they can operate and make decisions cognizant of all moving objects. In our previous research work we have developed a vision-based framework which is capable of detecting, tracking, and predicting the trajectories of people. These abilities constitute the core of a system capable of monitoring any developments concerning pedestrian motion. In this project, we aim at positioning our research for deployment. We plan to enhance our algorithms to make them suitable for operation in real traffic situations, in terms of accuracy and robustness against adverse operating conditions. Additionally, we will address the difficulties of bringing a new site into effective action, such as camera calibration and the initial identification of a context to make predictions. The expected outcome is an implementation of the enhanced framework which can be deployed in the field, and is capable of providing pedestrian information to Surtrac1—a real time traffic signal control system. The milestones envisioned, to accomplish in two years, are: 1) Development of enhanced approach for pedestrian detection; 2) Implementation of prototype suitable for operation in real time; and 3) Testing and characterization of prototype in the field. Our team is composed of a Master's student, and researchers with many years of experience in computer vision for autonomous systems.
This project addresses the need for timely and accurate information about pedestrian traffic in urban areas. This is particularly important at locations where it is not uncommon to find more pedestrians than vehicles during certain times of the day. Through the work proposed in this project, we move to position our research towards deploying a system for detecting, tracking, and forecasting human behavior in dynamic environments. This information can be used by other systems as part of an infrastructure-based framework to effectively protect the more vulnerable traffic participants. During recent efforts (Metro21), we have transitioned our software infrastructure from using LIDARs to video cameras as the primary sensor. We have developed a software framework where a static camera that has an unobstructed view of a road intersection is used to detect and track the motion of pedestrians. This represents the first level of awareness, i.e. how many people are present in the area, and their current location, as well as their speed and direction of motion. Furthermore, by learning from training data relationships between pedestrian paths and features in the environment, we can model explicit preferences in motion patterns, such as destinations and preferred routes. These models are used to predict the trajectories of pedestrians further in the future than what it is achievable using simple motion models. This constitutes the second level of awareness: the most likely paths to be traversed by the pedestrians. A sample output from our framework is shown in Fig. 1. Vision sensors are commonly used at traffic intersections: they have small space and power requirements, and their performance and robustness continues to improve as their cost becomes lower. They are suitable for widespread use. However, these sensors are more susceptible to extreme (i.e. too bright or too dim) or adverseillumination conditions (e.g. cast shadows induce color and illumination discontinuities in the image). Moreover, the accurate and robust detection of pedestrians from images is more challenging as the distance from the camera increases, or when cameras have wider fields of view (as is frequently the case with traffic cameras), which make objects on the scene to appear smaller in size. To address these issues, work still needs to be done to integrate our current algorithms in a way that its successful deployment is feasible. Hence, the goal of this project is the design and implementation of a software framework that processes video in real time to detect, track, and predict the trajectories of pedestrians at the site; that is robust and accurate over a wide range of operating conditions; that is easily installed and calibrated at every site; that is easy to maintain and diagnose on-site; and that can communicate the detection, tracking and prediction results to other users. To accomplish this goal, we will leverage different elements developed in other projects. We will use algorithms designed to detect poor or adverse illumination conditions, and to compensate for them to support pedestrian detection. We will also apply methods to extract priors from external sources of information, which can provide a richer initial context for predictions at a particular site. Additionally, we will adapt methodologies for intrinsic and extrinsic camera calibration for use at the site of each traffic camera. Finally, we will leverage our recent work on modeling interactions between pedestrians to extend our prediction framework, originally capable of producing predictions for each mover independently of other pedestrians, to generate multi-agent predictions using simple heuristics.
The work is planned over a 24-month time frame.
To understand and solve the issues involved in the deployment of our technology, and to evaluate the beneficial effects of pedestrian-related information, we have approached Rapid Flow Technologies, creators of Surtrac. We will collaborate with them in several ways: obtain selected examples of data for analysis; discuss the issues involved with the deployment of your system using actual traffic infrastructure;collaborate with the design of an interface with Surtrac; and receive feedback on the usefulness and practicalities of our system.
By the end of the project we expect to have a functional implementation of the enhanced algorithmic framework for pedestrian detection, tracking, and prediction, which is also capable of publishing information. This will enable us to run reliability tests and to carry out experiments to evaluate its performance, and to study and evaluate its usefulness by other applications. As for evaluation metrics, the primary objective is to detect the presence and behavior of people accurately and robustly in real time. The pedestrian detection accuracy will be measured using the percentage of false and missed detections over time. The accuracy of the predictions will be assessed by comparing the probabilistic distribution over paths generated by our models with the actual paths demonstrated by pedestrians. The secondary objective is to provide information to other systems (e.g. SURTRAC, driver assistant systems) to enhance their performance. Therefore, we will compare each system’s performance measures with and without our application to evaluate and quantify the improvements.
Name | Affiliation | Role | Position | |
---|---|---|---|---|
hebert@cs.cmu.edu | Hebert, Martial | Robotics Institute | Co-PI | Faculty - Tenured |
lenscmu@ri.cmu.edu | Navarro-Serment , Luis E. | Robotics Institute | PI | Faculty - Research/Systems |
Type | Name | Uploaded |
---|---|---|
Final Report | 065_finalReport.pdf | Aug. 17, 2018, 4:04 a.m. |
No match sources!
No partners!