Citywide reconstruction of cross-sectional traffic flow from moving camera videos

Vehicle re-identification and trajectory reconstruction using multiple moving cameras in the CARLA driving simulator

“…C. Vehicle detection and tracking 1) Vehicle detection using YOLOv7: For the detection of vehicles inside the CARLA driving simulator, we use the vehicle orientation dataset [4] to train a YOLOv7 model [8] to detect both vehicle class and orientation. The main reason to choose YOLOv7 to train the vehicle detection neural network is that YOLOv7 is significantly lightweight (75% reduction in parameters with 1.5% higher AP for the same base model) compared to its real predecessor YOLOv4 [13], [14] while achieving improved benchmark results on the COCO dataset [15].…”

“…We use the base model of YOLOv7 with pre-trained weights on the COCO dataset with an input size of 640×640. We first train with the real-world images from the vehicle orientation dataset [4] for 100 epochs with a learning rate of 0.001 on four Tesla A100 GPUs [16] and then use the synthetic vehicle orientation dataset [6] to fine-tune the next ten epochs with a reduced learning rate of 0.0001 to prevent large changes in the parameters. It should be noted that the vehicle orientation dataset contains 15 classes of vehicles, while the synthetic vehicle orientation dataset has 12 classes of vehicles due to the absence of bus class; thus, bus front, bus back, and bus side classes are not present.…”

“…As tracker state of the original SORT algorithm considers only single object class, similar to the research conducted in [4], we modify the state of the tracker to include the detected vehicle class (η), as shown in Eq. 1.…”

“…In recent studies [4], [5], the authors propose a novel framework using vehicle detection, vehicle tracking, vehicle localization, and map-matching on the OpenStreetMap road network for estimating traffic flow parameters from moving cameras mounted on the dashboard of a car. The authors suggest using crowd-sourcing for citywide traffic flow estimation, making the method highly scalable.…”

“…extends the methodology introduced in [4] to estimate traffic flow in a driving simulator CARLA to verify the efficacy of the proposed algorithm [4] under a variety of environmental conditions.…”