Modern vehicles rely on networked electronic control units (ECUs) that utilize communication protocols such as the Controller Area Network (CAN), which is susceptible to security vulnerabilities. This paper presents an innovative approach called Temporal Pattern Image-based Intrusion Detection System (TPI-IDS), explicitly designed for CAN. The proposed TPI-IDS converts the interarrival times of consecutive CAN frames into grayscale temporal images. Then, these images are processed by a Convolutional Neural Network (CNN) to classify CAN traffic as normal or abnormal. The CNN model is optimized through quantization for resource-constrained embedded systems, achieving a balance between computational efficiency and intrusion detection accuracy. TPI-IDS achieved F1-scores over 0.99 across all evaluated datasets, including perfect scores (F1 = 1.00) on the Spark subset. Deployed on the STM32F746 Discovery board, the quantized model achieved an inference time of 8.243 ms, used 21.74 KiB RAM, and required 41.46 KiB of total flash memory. Moreover, the quantized TPI-IDS model has only 6.134 KB of parameters, making it one of the most lightweight approaches compared to existing methods and demonstrating strong potential for deployment in resource-constrained embedded automotive systems.