Lithium-ion batteries (LIBs) are widely used in electric vehicles (EVs) due to their long cycle life, high energy density, and fast charging abilities. However, the natural decline in LIB performance over time, leading to reduced capacity and power, presents challenges. This is especially true in battery packs, where inconsistencies among individual cells and non-uniform aging patterns complicate the estimation of state of health (SOH). Capacity serves as a crucial parameter for assessing SOH, providing insights into remaining energy storage capability and maintenance needs. Many studies have focused on estimating the capacity of single LIB cells or experimental datasets, but applying LIB packs in real-world situations requires customized models that consider environmental factors such as temperature and road conditions. This research introduces a novel hybrid model named CNN-Transformer for LIB pack capacity estimation, leveraging a combination of 1D-Convolutional Neural Network (CNN) feature extraction and transformer networks. The proposed method is evaluated against traditional recurrent neural network (RNN)-based methods, revealing superior performance. This evaluation highlights the effectiveness of the CNN-Transformer in managing long-term dependencies and overcoming the limitations of attention mechanisms found in transformer networks. Finally, the proposed methods are validated using real-world EV dataset, demonstrating practical effectiveness in LIB pack capacity estimation. In conclusion, this research contributes innovative approaches to real-world LIB pack capacity estimation, addressing key challenges and showcasing superior performance compared to traditional methods. These advancements hold significant promise for advancing the energy storage industry and ensuring the safe and stable operation of LIB packs in diverse applications.