Machine Learning-based automated systems are highly desirable for their superior accuracy in a dynamically varying application environment. Also, Processing-in-Memory (PIM) accelerators are suitable for low-power hardware-software co-design. In this work, we adopt an in-memory online learning approach to train CNN algorithms on uPIM, a custom-designed & ultra-low-power Look-up Table (LUT) based programmable PIM accelerator. Our performance-aware design is capable of achieving an impressive 72% and 83.4% accuracy on the German Traffic Sign Recognition Database (GTSRB) and the Belgian Traffic Sign Dataset (BTSD) respectively at up to 96.5 FPS of CNN inference throughput and 75x energy-efficiency over state-of-the-art GPUs.