This report explores the integration of high-end neural network models into embedded systems, its advantages over the high-end computers and Cloud-based inferencing to overcome the drawbacks such as cost, power consumption, latency, security concerns, and reliability. The STM32 microcontrollers are widely used in various embedded applications and are known for their high performance, low power consumption, and extensive connectivity options. A CNN architecture, ConvNet3, is proposed and fine-tuned for object classification and navigating obstacles in real-time scenarios, leveraging the limited resources of the STM32F401RE board. The performance of these models is evaluated, considering factors like prediction speed, accuracy, and model size. The findings highlight the trade-offs between accuracy and size reduction achieved through compression techniques. Overall, this research contributes to advancing CNN-based solutions for embedded systems and offers insights for model selection and data preparation, facilitating improved navigational capabilities and object recognition in resource-constrained environments.