In the scenario of large-scale automatic apple warehousing, the counting model based on computer vision must balance lightweight design with detection accuracy. Traditional models involve large parameters and high computational costs, making real-time operation challenging;moreover, in complex environments characterized by dense apples and severe occlusion, issues such as boundary blurriness and high false detection rates arise. To address these challenges, an improved CGW-YOLO v8 model was proposed. Firstly, by replacing the C2f module in the backbone network with the lightweight GhostNet module and incorporating a feature channel reweighting mechanism, the model's parameter count was significantly reduced. Secondly, the CSPHet module was employed, which utilized heterogeneous multibranch convolution and a dual-path feature fusion strategy to enhance the boundary distinguishing capability of densely packed apples while decreasing the number of parameters. Lastly, a loss function based on Wasserstein distance loss was adopted to replace the traditional IoU metric, effectively reducing the false detection rate in densely stacked scenarios. Experimental results indicated that the model's mean average precision (mAP@0.5) was improved to 95.8%, representing a 1 percentage points increase compared with that of the original model, with precision and recall increased by 1.1 percentage points and 1.3 percentage points, respectively, compared with that of the original model. Both parameter count and computational volume were decreased by 24.4% and 23.2% relative to that of the baseline model. To meet the dual requirements of real-time counting and accuracy in the inventory production process, the DeepSORT tracking algorithm was integrated to achieve continuous tracking and accurate counting of apples across video frames. A counting strategy based on trajectory management was designed, which counted only when the target first crossed a virtual counting line, effectively avoiding issues of duplicate statistics and missed counts. Experimental results demonstrated that the proposed improved method exhibited strong robustness and high counting accuracy, particularly in complex backgrounds, such as densely arranged apples and partially obscured scenes.