سیستم زمان واقعی برای نظارت دوچرخه سواران و عابران

Camera based systems are routinely used for monitoring highway traffic, supplementing inductive loops and microwave sensors employed for counting purposes. These techniques achieve very good counting accuracy and are capable of discriminating trucks and cars. However, pedestrians and cyclists are mostly counted manually. In this paper, we describe a new camera based automatic system that utilizes Kalman filtering in tracking and Learning Vector Quantization for classifying the observations to pedestrians and cyclists. Both the requirements for such systems and the algorithms used are described. The tests performed show that the system achieves around 80–90% accuracy in counting and classification.

The analysis of traffic flows has traditionally been based on inductive and microwave sensors. These technologies are not, however, capable of reliable monitoring of pedestrians and cyclists for two reasons. First, the reliability of inductive sensors depend on the materials in the vehicle, and second, the structure of pedestrian traffic is weak and does not focus on easily recognizable lanes. It is necessary to find an alternative technique for pedestrian traffic analysis due to the problems of the traditional approaches. Traffic flow statistics are mainly used by road administration which tries to resolve the needs for new pavements, cycle paths and tunnels, for instance. Currently, such surveys are performed manually. However, manual counting is tedious and typically continues only for a short period of time which makes the results quite unreliable. A natural choice for monitoring pedestrians and cyclists automatically is to use visual perception. However, there are still quite few camera based systems developed for people detection and tracking. A well known undertaking in this area was the ESPRIT project PASSWORDS [2]. Another visual surveillance application is W4 [6], which is a real time system for determining interactions between people. Most of the systems presented are based on an assumption of a stationary camera which greatly simplifies the problem of human detection, although solutions for pedestrian detection from moving platforms have also been suggested, for example, in Refs. [5] and [10]. Furthermore, a model based approach for pedestrian detection was proposed in Ref. [11], and a statistical framework for learning and recognizing human movements in Ref. [3]. In human tracking, there are many visual clues that can be utilized. These include color histograms [4], coherent connected regions or ‘blobs’ [12], and object contours [1]. Kalman filtering is probably the most commonly used algorithm for implementing the tracker, although recently Condensation algorithm [7] and mean shift algorithm [4] have shown to provide certain advantages especially in the presence of significant background clutter. Despite of the methodological advancements, many of the solutions proposed are computationally expensive and they make limiting assumptions about the target appearance and dynamics. Moreover, there is often a long way to go from a controlled laboratory environment to unpredictable outdoor scenes. When the project for developing an automatic traffic monitoring system was started in 1994, there were no commercial systems available for that purpose. According to our knowledge there still exist no other systems capable of fulfilling even partly the following requirements that were specified by our user group: • counting of the volume of the traffic • classification of the traffic into separate traffic classes (pedestrians, cyclists,…) • route tracing • counting accuracy better than 90% • classification accuracy better than 85% • portability and easy installation • one week stand alone operation time • transfer of data into a portable field terminal through a telephone network or via a memory card In addition, the weather conditions set extra requirements on the hardware. The device should work in various temperatures (from −20 to+30 °C), in rain, snowfall, and wind. Also, adaptation to a wide illumination range with sudden changes is typically very difficult for camera based systems. One of the main problems in developing a stand alone system is the power consumption. The limited capacity of the batteries and long operation time allows us to use only economic hardware solutions at the cost of processing speed. This also forced us to divide the monitoring problem into two parts: real-time tracking and off-line analysis. The tracking unit is installed beside the road on a pole. It performs several image processing tasks to detect and track moving objects in the scene. The shape and the trajectory of the objects are stored in a removable memory card to wait for further processing. After several days of stand alone operation, the memory card is fetched and the data is transferred to a separate desktop computer equipped with analysis software. The objects are then classified to pedestrians and cyclists based on the features stored during the tracking phase. The architecture of the monitoring system is depicted in Fig. 1.In Section 2, we will describe the basic techniques used in the tracking unit including the image processing steps for feature extraction and the Kalman filter based tracking scheme. Section 3 introduces the Learning Vector Quantization (LVQ) based analysis for producing the final counting and classification result. The preliminary test results are given in Section 4.

In this paper, we presented a new approach to count and classify cyclists and pedestrians automatically. The test results are promising, but there is still room for improvement. Especially, object recognition and classification could be enhanced by using more descriptive shape features. The QTC codes are fast to extract, but they are very sensitive to noise. Using correlation masks or higher order statistics could provide a better solution, but they also increase the amount of computation needed. Another problem is the shadows that occasionally remain connected to the silhouette after motion detection. They can cause the tracker to lose the target resulting in disjointed trajectories. They also cause the feature vectors to be scattered within the same category, which makes the classification task more difficult. Dividing the system into real-time tracking unit and off-line analysis seems to be a good solution. We believe that in the future this architecture will be a basis for similar monitoring systems.