Intralogistics (or internal logistics) focuses on the management and optimization of internal production and distribution processes within warehouses, distribution centers, and factories. Automated handling systems play a crucial role in the internal logistics of several industries such as e-commerce, postal messaging, retail, manufacturing, airport transport, etc. These systems are composed by multiple high-speed conveyor lines that provide safe and reliable transportation of a large volume of goods and merchandises while reducing costs.

The automation of the conveying process relies on the identification and the real-time tracking of the transported objects. In this thesis, we designed a tracking solution that employs a network of smart cameras with an overlapping field of view. The goal is to provide a real time tracking information about conveyed objects to control the automated handling system.

Multiple object tracking is a fundamental problem of computer vision that has many applications such as video surveillance, robotics, autonomous cars, etc. We integrated several building blocks traditionally applied to traffic surveillance or human activities monitoring to constitute a tracking pipeline. We used this baseline tracking pipeline to characterize contextual scene information proper to the conveying scenario. We integrated this contextual information to the tracking pipeline to enhance its performance. In particular, we took into account the state of moving objects that become stationary in the background subtraction step to prevent their absorption to the background model. We have also exploited the regularity of objects’ trajectory to enhance the motion model associated with the tracked objects. Finally, we integrated the precedence ordering constraint among the conveyed object to reidentify them when they are close to each other.

We have also tackled practical problems related to the optimization of the execution of the proposed tracking pipeline in the multi-core architectures of smart cameras. In particular, we proposed a dynamic learning process that extracts the region of the image that corresponds to the conveyor lines. We reduced the number of the processed pixels by restricting the processing to this region of interest. We also proposed a parallelization strategy that adaptively partitions this region of interest of the image, in order to balance the workload between the different cores of the smart cameras.

Finally, we proposed a multiple cameras tracking algorithm based on event composition. This approach fuses the local tracking generated by the smart cameras to form global object 

trajectories. It also includes information from third party systems such as the object destination entered by human operators.

We deployed and validated the proposed approach for the control of a sorting system in a postal distribution warehouse. A network of 32 cameras tracks more than 400.000 parcel/day on the injections lines feeding the sorting system. The tracking error rate is less than 1 parcel in 1000 (0.1%).