Objective Urban parks play a vital role in enhancing residents’ physical and mental well-being and offering leisure opportunities. Their vitality has become a crucial indicator of urban spatial quality and public welfare. Rapid urbanization has further intensified the imbalance in the allocation of public service resources. Existing research, which primarily relies on heat maps, mobile signaling data, or ground-based camera monitoring, can reveal macroscopic trends but fail to capture the dynamic spatiotemporal characteristics of crowd distribution at the micro scale. Meanwhile, aerial photography obtained through unmanned aerial vehicle (UAV) offers high spatial resolution and flexible data acquisition capabilities, while the advancement of object detection algorithms based on deep learning presents new technological opportunities for crowd recognition in complex urban environments. This research aims to develop and validate a micro-scale vitality measurement method for urban parks based on aerial time-series imagery and an improved object detection model. The method seeks to reveal the spatiotemporal patterns of crowd distribution, identify high-frequency vitality nodes and their driving mechanisms, and provide data support and strategic insights for optimizing the spatial layout, facility allocation, and refined management of parks. Taking Xi’an Xingfu Linear Park as an example, the research focuses on analyzing vitality intensity, fluctuation, and spatial balance at a fine spatiotemporal scale.

Methods Between March 27 and 30, 2025, continuous UAV-based aerial photography was conducted at a fixed altitude of 75 m during six standard time periods (08:00, 10:00, 12:00, 14:00, 16:00, 18:00), yielding over 2,300 high-resolution images. A manually annotated dataset of 2,000 sub-images with 12,340 pedestrian instances is constructed for model training. To address challenges of small-scale targets and complex occlusions in aerial imagery, an enhanced YOLO11m-CBAM model is developed by embedding a convolutional block attention module (CBAM) into YOLO11m. The improved model achieves notable performance gains: mAP50 increases from 77.1% to 81.3%, mAP50–95 from 45.6% to 51.7%, with precision and recall reaching 86.4% and 72.0% respectively, demonstrating enhanced robustness under medium and low occlusion conditions. Detection outputs are orthorectified to geographic coordinates to construct a structured spatiotemporal dataset. Spatial analysis employs kernel density estimation, coefficient of variation (CV), spatial Gini coefficient, and the “latitude-population” curve to characterize multidimensional vitality patterns.

Results The temporal analysis results indicate that the overall utilization of Xingfu Forest Belt exhibits a distinct “dual-peak” pattern. On rest days, the number of visitors reaches 2,112 at 10:00 and 3,641 at 16:00, reflecting typical peaks of family and leisure activities. The daily coefficient of variation (CV = 38.67%) is relatively low, suggesting stable visiting patterns with activity concentrated in leisure hours. In contrast, on working days, vitality peaks occur at 10:00 and 18:00, corresponding to post-commuting and after-work relaxation periods, respectively. The higher daily visiting (CV = 55.34%) indicates a more uneven temporal distribution of activities. Notably, 12:00 represents the lowest point of visiting (the minimum number of visitors is only 595, and the average number is 883), implying underutilization of space during midday and suggesting potential opportunities for future facility optimization or time-specific programming. The spatial equilibrium analysis further reveals that during peak hours (14:00 and 16:00), the spatial Gini coefficient reaches 0.44 – 0.48, indicating a strong concentration of vitality in specific functional zones and a pronounced spatial polarization effect. In contrast, the Gini coefficient drops to 0.24 during off-peak periods (08:00 and 12:00), reflecting a more dispersed and evenly distributed use of space. At 18:00, the Gini coefficient remains between 0.38 and 0.41, suggesting a moderate level of aggregation in the evening. Overall, the vitality of Xingfu Forest Belt demonstrates a dynamic pattern of “daytime polarization with evening recovery”. In terms of spatial distribution, vitality hotspots are primarily concentrated along the central and northern segments of the belt, forming localized peaks. The emergence of these core areas is driven by two main factors: 1) the attraction of fixed functional facilities such as children’s play areas, fitness zones, and square-dancing spaces; and 2) the temporal aggregation generated by periodic activities, including weekend family events and morning exercise. At the macro scale, the concentration of residential and educational land uses, high accessibility to bus stops, and the scarcity of comparable recreational facilities jointly reinforce the sustained vitality of the central children’s play area. Maintaining consistently high footfall and strong spatial spillover effects across multiple time periods, this area serves as a key vitality hub within the overall spatial structure of Xingfu Forest Belt.

Conclusion The research demonstrates that the proposed UAV-based and YOLO-based vitality measurement framework provides high spatiotemporal resolution at the micro-park scale, enabling accurate identification of vitality hotspots, temporal fluctuations, and spatial imbalances. This approach offers an operational, quantitative basis for optimizing facility layouts, designing flexible spaces, and implementing differentiated management strategies. Methodological limitations are also discussed: The approach performs reliably in spring, autumn, and winter with low to moderate vegetation coverage, but may encounter partial omissions under dense canopy or multi-layer pergola structures in summer. To enhance applicability, future improvements include multi-drone and multi-view data acquisition, infrared thermal imaging to mitigate occlusion, air-ground data fusion, inter-frame trajectory matching to distinguish stay/pass behaviors, and fine-grained activity recognition. Overall, the proposed method provides a replicable technical pathway and empirical reference for refined park governance and smart park development. The findings contribute to advancing quantitative urban vitality assessment and provide methodological insights for integrating AI and spatial analysis in urban landscape research.