Autonomous object detection dataset: a study on Bangladeshi roadways
Abstract
This research aims to enhance object detection in autonomous driving by testing
the three new advanced YOLO models, namely, YOLOv5, YOLOv7, and YOLOv8,
on a private dataset proposed specifically for Bangladesh. It is suggested that
the dataset include conventional vehicles such as cars, trucks, and buses, nonconventional
modes of transport such as rickshaws and CNGs, and complex scenes
with poor visibility or high congestion. The contribution of this research involves
the creation of a custom dataset that overcomes the deficiencies in other publicly
available datasets in terms of improved representation for unconventional vehicles
and managing different traffic and weather conditions. YOLOv5, YOLOv7, and
YOLOv8 are trained and tested by comparing performances concerning mean Average
Precision (mAP), precision, recall, and inference speed. YOLOv8 was the
best performer in different difficult conditions related to traffic and even challenge
lighting environments, while YOLOv5 was observed to perform more efficiently in
the case of real-time applications where the resources are constrained. YOLOv7
strikes a good balance between speed and accuracy, hence is suitable for moderately
complex environments. However, all the models were vulnerable to class imbalance
and the detection of smaller or occluded objects. Comparing these models with
their predecessors, like Faster R-CNN, testifies that the YOLO models perform the
best in real time among other models. In general, but especially for YOLOv8, more
improvements should be made on dataset representation, occlusion handling, and
edge optimization if it is to find a place in autonomous driving systems in the near
future. Hence, the research relies on a custom-built private dataset, which effectively
captures the dynamic and unstructured environment of traffic in Bangladesh
in much finer detail compared to other, earlier datasets.