Nowadays, all efforts in traffic management systems are focused on finding the best solution for detection and trajectory reconstruction using AI techniques. Many companies are engaged to improve deep learning models to build algorithms more and more efficient and able to better generalize.
Traffic managers want a system capable of recognizing each type of vehicle, behaviour and anomaly, in a way that is not affected by weather, light or environmental conditions.
No matter the type of object that these different analytics are being trained to recognize, there is always an approach that is based on the analysis of a single camera, treating them independently one to each other. Data collected by these cameras is isolated. One camera can only monitor a fixed view and there is no bridge to share the monitoring information with each other. Data coming from a previous observation will not be propagated through the next element of the chain of computation. This information will be lost.
The role of managers and operators
The intra-camera correlation is usually carried out by traffic managers and operators manually, with just the aid of camera names, numbers and kilometers. Some other times a certain degree of correlation is obtained by applying logical rules between cameras (i.e. suppressing alarms from the following camera if they were already detected by the previous one).
Exploiting an unlimited potential
We think that the future will be considering the surveillance system as a whole. This information has to be used because it has a valuable relevance in order to simplify, accelerate, focus and pre-alert operators and traffic managers.
Due to the exponential growth of IP video surveillance cameras, the opportunity to take advantage of the rich information from the multi-camera systems is immense.Following the need to represent the network of connected cameras as a single entity, where vehicles, bicycles and pedestrians can be monitored from when they enter the scene until they exit, Sprinx has developed the Multi-camera tracking module.
The Multi-camera tracking module is capable of identifying and reconstructing the trajectory of all the objects moving in the environment over the field of view of all the cameras. Doing this also allows the detection of long-term vehicle behaviours and making inferences during all their journey.Furthermore, information from previous observations can help in reducing noise in measurements, removing ambiguities in detection and improving tracking performance by focusing only on specific regions of interest.
Unlike the estimation of traffic information based on single-camera, working with multi-camera systems is much more challenging but, as previously outlined, is highly accurate. The Multi-camera tracking module can overcome real-world scenarios, extracting features of the same object from several camera views, even if orientation and lighting conditions can create a massive variation in its camera projected aspect.
Spatial-temporal constraints are also integrated by the Multi-camera tracking module to correctly track the objects in the environment as they pass through the network of cameras. The system can estimate the relative position of the road lanes and surfaces (in terms of the full reconstruction of different positions and orientations) between the different cameras to compute on which camera an object shall appear when exiting from another one. This allows the creation of graphs of connected lanes.