Pedestrian counting in scenic spots is an important technical means to promote intelligent scenic spot management, which has far-reaching impacts such as enhancing visitor experience, optimising resource allocation and guaranteeing safety. Aiming at the limitations of existing pedestrian counting methods in terms of accuracy and real-time performance, this paper proposes a lightweight pedestrian counting method based on improved YOLOv10n and DeepSort. In the detection stage, DC-FFA, an attentional feature fusion structure based on lightweight dynamic convolution, is designed to effectively retain the shallow information and enhance the model representation ability; SimSPPF module with a simplified structure is introduced to reduce the computation amount and improve the efficiency; and the Concat module of FPN is combined with BiFPN to form the Concat_BiFPN module, which promotes the bi-directional flow of features between different scales. Bidirectional flow. In the tracking phase, a new pedestrian re-identification network SEAMNet36 is proposed to enhance the differentiation ability of similar pedestrians, and Powerful-IOU is used to replace the traditional IOU matching strategy to improve the accuracy and robustness of target matching. In the counting phase, a counting algorithm based on biconvex trackline measurement is used to derive the number of entering and exiting pedestrians. The experimental results show that the improved YOLOv10n improves 1.3% compared to the original model mAP50, the improved DeepSort algorithm improves 5.5% compared to the original model MOTA, and the MOTP improves 3%, which is better than the traditional method in terms of accuracy, real-time and computational resource consumption, and has a good application prospect.