A cyclist-centric 360^° panoramic dataset for safety-critical object detection in real-world cycling scenarios

Cycling safety is becoming an increasingly significant challenge for urban transportation, and vision-based object detection technology for cycling scenes has been widely studied. Datasets are an essential component of studying object detection in cycling scenes. However, existing datasets often have limitations, including limited scene diversity and incomplete coverage of safety-critical objects. What is particularly critical is that the data acquisition perspective is not from the perspective of real cyclists, and the limited field of view cannot cover blind spot risks. To address these limitations, we introduce PanoCycle360, a novel 360^° panoramic dataset designed for complex cycling scenarios. The dataset is collected using a 360^° panoramic camera mounted on the cyclist's helmet, providing full 360^° coverage to eliminate blind spots in cycling and encompassing all common safety-critical objects and real-world cycling scenarios. The dataset comprises 10,055 panoramic images that underwent manual annotation, covering nine high-risk classes, such as E-Bike Rider, Person, Car, Van, Bus, Truck, Cyclist, Cargo Tricycle, and Auto Rickshaw, resulting in 102,171 bounding boxes. We evaluated PanoCycle360's applicability in single-stage, two-stage, and Transformer-based object detection frameworks across various algorithms with different parameter scales. In combination with the experimental results, it can be concluded that PanoCycle360 enables reliable evaluation in multiple scenarios. The development of the PanoCycle360 dataset holds significant implications for advancing cyclist-centric safety research and developing safety-critical object detection systems for real-world cycling scenarios worldwide.