The observation and monitoring of traffic with smart visions systems for the purpose of improving traffic safety has a big potential. Today the automated analysis of traffic situations is still in its infancy--the patterns of vehicle motion and pedestrian flow in an urban environment are too complex to be fully captured and interpreted by a vision system. 3In this work we present steps towards a visual monitoring system which is designed to detect potentially dangerous traffic situations around a pedestrian crossing at a street intersection. The camera system is specifically designed to detect incidents in which the interaction of pedestrians and vehicles might develop into safety critical encounters. The proposed system has been field-tested at a real pedestrian crossing in the City of Vienna for the duration of one year. It consists of a cluster of 3 smart cameras, each of which is built from a very compact PC hardware system in a weatherproof housing. Two cameras run vehicle detection and tracking software, one camera runs a pedestrian detection and tracking module based on the HOG dectection principle. All 3 cameras use sparse optical flow computation in a low-resolution video stream in order to estimate the motion path and speed of objects. Geometric calibration of the cameras allows us to estimate the real-world co-ordinates of detected objects and to link the cameras together into one common reference system. This work describes the foundation for all the different object detection modalities (pedestrians, vehicles), and explains the system setup, tis design, and evaluation results which we have achieved so far.
The observation and monitoring of traffic witih smart vision systems for the purpose of improving traffic safety has a big potential. Embedded loop sensors can detect and count passing vehicles, radar can measure speed and presence of vehicles, and embedded vision systems or stationary camera systems can count vehicles and estimate the state of traffic along the road. This work presents a vision system which is targeted at detecting and reporting incidents at unsecured railways crossings. These crossings, even when guarded by automated barriers, pose a threat to drivers day and night. Our system is designed to detect and record vehicles which pass over the railway crossing by means of real-time motion analysis after the red light has been activated. We implement sparse optical flow in conjunction with motion clustering in order to detect critical events. We describe some modifications of the original Lucas Kanade optical flow method which makes our implementation faster and more robust compared to the original concept. In addition, the results of our optical flow method are compared with a HOG based vehicle detector which has been implemented and tested as an alternative methodology. The embedded system which is used for detection consists of a smart camera which observes one street lane as* well as the red light at the crossing. The camera is triggered by an electrical signal from the railway as soon ss a vehicle moves over th this line, image sequences are recorded and stored onboard the device.
The observation and monitoring of traffic with smart visions systems for the purpose of improving traffic safety has a
big potential. For example embedded vision systems built into vehicles can be used as early warning systems, or
stationary camera systems can modify the switching frequency of signals at intersections. Today the automated analysis
of traffic situations is still in its infancy - the patterns of vehicle motion and pedestrian flow in an urban environment are
too complex to be fully understood by a vision system.
We present steps towards such a traffic monitoring system which is designed to detect potentially dangerous traffic
situations, especially incidents in which the interaction of pedestrians and vehicles might develop into safety critical
encounters. The proposed system is field-tested at a real pedestrian crossing in the City of Vienna for the duration of one
year. It consists of a cluster of 3 smart cameras, each of which is built from a very compact PC hardware system in an
outdoor capable housing.
Two cameras run vehicle detection software including license plate detection and recognition, one camera runs a
complex pedestrian detection and tracking module based on the HOG detection principle. As a supplement, all 3
cameras use additional optical flow computation in a low-resolution video stream in order to estimate the motion path
and speed of objects. This work describes the foundation for all 3 different object detection modalities (pedestrians,
vehi1cles, license plates), and explains the system setup and its design.
This paper presents a car detection system that is able to work in close to real-time on a smart camera. A cascade of
histograms of oriented gradients was used as a detector. The algorithm and code were optimized for speed to meet the
real-time constraints, without loosing too much on detection quality. The system is now able to process 10 frames per
second on an Atom Z530 (1.6 GHz) processor used in the smart camera. The application on which the paper is based is
ready to detect cars in real world scenarios. It is planned to extend it to also track and analyze the driver behavior
Proc. SPIE. 6764, Intelligent Robots and Computer Vision XXV: Algorithms, Techniques, and Active Vision
KEYWORDS: Digital signal processing, Image compression, Detection and tracking algorithms, Cameras, Image processing, Data processing, Signal processing, Image filtering, Optical flow, Computer arithmetic
This work presents the implementation of the Kanade-Lucas-Tomasi tracking algorithm on a Digital Signal Processor
with a 40-bit fixed-point Arithmetic Logic Unit built into a smart camera. The main goal of this work was to obtain realtime
frame processing performance while loosing as little tracking accuracy as possible. This task was motivated by
increasing demand for the application of smart cameras as main data processing units in large surveillance systems,
where factors like cost and demand of space are excluding PCs from this role.
In a first effort the modification of the Kanade-Lucas-Tomasi to integer numbers was performed and then in the next
step the influence on stability and accuracy of this modification was investigated. It is demonstrated how changing the
numeric data type of intermediate results within the algorithm from float to integer, and decreasing the number of bits
used to store variables, affects tracking accuracy. Nevertheless the DSP implementation can be used where the
computation of optical flow based on a tracking algorithm needs to be done in real-time on an embedded platform where
limited subpixel accuracy can be tolerated. As a further result of this implementation we can conclude that a DSP with a
fixed-point arithmetic logic unit can be very effectively applied for complex computer vision tasks and is able deliver
good performance even compared to high-end PC architectures.
This paper describes the implementation of a pedestrian detection system which is based on the Histogram of Oriented
Gradients (HOG) principle and which tries to improve the overall detection performance by combining several part
based detectors in a simple voting scheme. The HOG feature based part detectors are specifically trained for head, head-left,
head-right, and left/right sides of people, assuming that these parts should be recognized even in very crowded
environments like busy public transportation platforms. The part detectors are trained on the INRIA people image
database using a polynomial Support Vector Machine. Experiments are undertaken with completely different test
samples which have been extracted from two imaging campaigns in an outdoor setup and in an underground station. Our
results demonstrate that the performance of pedestrian detection degrades drastically in very crowded scenes, but that
through the combination of part detectors a gain in robustness and detection rate can be achieved at least for classifier
settings which yield very low false positive rates.
State of the art algorithms for people or vehicle detection should not only be accurate in terms of detection performance and low false alarm rate, but also fast enough for real time applications. Accurate algorithms are usually very complex and tend to have a lot of calculated features to be used or parameters available for adjustments. So one big goal is to decrease the amount of necessary features used for object detection while increasing the speed of the algorithm and overall performance by finding an optimum set of classifier variables. In this paper we describe algorithms for feature selection, parameter optimisation and pattern matching especially for the task of pedestrian detection based on Histograms of Oriented Gradients and Support Vector Machine classifiers. Shape features were derived with the Histogram of Oriented Gradients algorithm which resulted in a feature vector of 6318 elements. To decrease computation time to an acceptable limit for real-time detection we reduced the full feature vector to sizes of 1000, 500, 300, 200, and 160 elements with a genetic feature selection method. With the remaining features a Support Vector Machine classifier was build and its classification parameters further optimized to result in less support vectors for further improvements in processing speed. This paper compares the classification performance, of the different SVM's on real videos (some sample images), visualizes the chosen features (which histogram bins on which location in the image search feature) and analyses the performance of the final system with respect to execution time and frame rate.
This paper describes a close to real-time scale invariant implementation of a pedestrian detector system which is based on the Histogram of Oriented Gradients (HOG) principle. Salient HOG features are first selected from a manually created very large database of samples with an evolutionary optimization procedure that directly trains a polynomial Support Vector Machine (SVM). Real-time operation is achieved by a cascaded 2-step classifier which uses first a very fast linear SVM (with the same features as the polynomial SVM) to reject most of the irrelevant detections and then computes the decision function with a polynomial SVM on the remaining set of candidate detections. Scale invariance is achieved by running the detector of constant size on scaled versions of the original input images and by clustering the results over all resolutions. The pedestrian detection system has been implemented in two versions: i) fully body detection, and ii) upper body only detection. The latter is especially suited for very busy and crowded scenarios. On a state-of-the-art PC it is able to run at a frequency of 8 - 20 frames/sec.