To improve safety, mitigate traffic congestion and to reduce pollution caused by vehicles, infrastructure-side sensors can be used, especially at critical locations in cities. In the future, infrastructural safeguarding has large potential, due to availability of advanced sensors (camera, radar, lidar) and vehicle-to-infrastructure(V2I) communications. Currently, camera-based systems are widely used to monitor traffic violations. A smart combination of multiple sensors like camera-RADAR or camera-LIDAR is used to determine the precise velocity and position of the traffic participants. In such scenarios RADAR/LIDAR will be responsible for detection of velocity or position and cameras will be used to identify the traffic participants, i.e. for object classification. However, processing of large amount of data is necessary at the sensor nodes. With the evolution of technology and availability of higher computational power, such systems will become affordable and smarter. Additional hardware can enable such systems to communicate with other traffic participants in order to increase safety and efficiency. Additional hardware and computational power will be limited due to cost overhead, size, weather conditions and power consumption limitations in the open-air roads. To mitigate such limitations, we have could-based solutions where data are acquired at the road side units but processed remotely in the cloud. Although it is a valid solution, it brings limitation regarding the required high bandwidth and is also a potential threat for data leaks, e.g. privacy and data security. To have a large detection range a camera imager needs to have a large chip area and high number of pixels. Therefore, the image size gets large even if the large number of pixels is not required for objects in short distance. In this paper an image pre-processing method is developed to reduce the sensor data size, which in turn reduces the computational power to process or the bandwidth to transmit the data. An increase of detection range is possible keeping the data size at an acceptable level. Reducing the sensor data size is a benefit and reduces the dependency of cloud-based solutions. Even in case of using a cloud-based solution, reduced data size will result in a lower network load, that increase overall performance of could base systems. In the paper, YOLO-V3 is used for object detection and classification of traffic participants. In Addition, the fixed installation of the camera in the infrastructure allows to apply methods for depth estimation when using only mono cameras. The improvement and accuracy of the depth estimation is benchmarked using data from RADAR and LiDAR sensors as ground truth, which are installed at the same sensor node as the camera, i.e. the data of Radar and LiDAR are fused to the camera data.
Mr. Numan Senel, Technische Hochschule Ingolstadt, GERMANY Mr. Shrivatsa Udupa, CARISSMA Institute of Electric, Connected and Secure Mobility, TH Ingolstadt, GERMANY Prof. Dr. Gordon Elger, Fraunhofer IVI – Applied Research Center for Connected Mobility and Infrastructure VMI, GERMANY