Is machine vision (machine that can see) required in future vehicles? Are cameras useful in road traffic optimization? What kind of artificial intelligence does a smart road infrastructure assume? 

Many cars already have road sensors installed on board. Most common feature is the detection and real-time recognition of common traffic signs with a device built into a moving vehicle. We already have mature technology for this. Today Traffic Sign Recognition (TSR) is widely used to detect and recognize the traffic signs passed and present the driver's appropriate information. In general, no further actions are taken in the vehicle. In some cases, the TSR system not only detects and recognizes road signs. TSR has the right to trigger limited actions like speed reduction. In level 5 of automation, TSR should provide practically 100% reliability, and what is more, it should detect not only internationally standardized signs but anything that can seriously influence the driving of the autonomous vehicle.

For the robust and reliable operation of CCAM traffic systems, the road infrastructure must also be equipped with sensors capable of identifying vehicles in motion and foreign objects, endangering traffic safety. 

The primary input to such a traffic data gathering system is a live video stream captured by cameras located over the road lanes. (Other input, like noise analysis, could be an added value in accuracy and reliability.) 

The design of such systems is still in the research phase. We have little experience with what objects need to be sensed or what sensitivity parameters provide a safe and economical optimum. Examples of such parameters are resolution, depth of field, detection angle, and distance. Camera toll control systems are good practice for applying camera picture content recognition with adequate resolution and noise tolerance. 

However, increasingly high-level automated intelligent traffic management requires the collection of more and more detailed information. Solving the problem is not trivial. Vehicles traveling at significant speeds on the road can be useful to observe not only individually but also in groups. They form a casual group whose members influence each other. Controlling near-optimal yet safe vehicle traffic will be a complex task when many highly automated vehicles travel. Little is known about this complex system's behavior in which the cooperative operation of autonomous individuals can be mutually beneficial. Evolution has developed many examples of beneficial swarm behavior in the animal kingdom. An example is the movement of flocks of birds. 

Engineering researchers need to investigate, among other tasks, what properties camera images are needed for automation of road transport, and how to ensure their effective analysis, communication, and use. Experiments with different placement of the cameras are also worth pursuing. An analysis is inconceivable without intelligent IT systems. At present, the application of artificial intelligence, including machine learning, seems promising. It may be useful to extract morphological image processing, image histograms, edge detection, and other image features. For effective content-based analysis of video streams, labeling them is practical, but mass implementation is slow and costly. It is worth researching the latest unsupervised learning methods. Traffic management research requires the recording of large amounts of real traffic data. Data cleansing, segmentation, object recognition, and classification procedures and anonymization to protect personal data will be required. The analysis of multimedia content using data mining terminology can also be called multimedia mining. Artificial intelligence (including machine learning) has a common field, shared features with multimedia mining. In both cases, the search for patterns, correlations, and deeper relationships within large data sets is the goal for predicting outcomes. Machine learning algorithms are data-driven, so they deal with large amounts of data, and their great advantage is their ability to learn the regularities inherent in the data flexibly. 

In addition to direct analyzes, it is worthwhile for simulation experiments to derive abstract data sets from real data that retain their statistical characteristics. The topic of object recognition includes several interrelated issues, including object localization: the purpose is to determine an object's position that can be identified in the input image. Object detection means the location of all objects in the input image. Object identification is for behavioral research purposes, it is necessary to assign a unique identifier to each object and detect its all occurrences. Object identification is not required for direct traffic management purpose, but may result useful and flexible patterns. A general vehicle object recognition system's operation can be divided into three tasks: vehicle detection, extraction of characteristics (vehicle category) and unique identification if necessary and permitted (for research and development purposes). The first two tasks, object detection and feature extraction can be performed simultaneously.

‍

Written by Data Management and Communication Team