Deep Learning-Enabled Squirrel Detecting Prototype Development and Processing Analysis

In this paper, the authors present the development of an object detection prototype aimed at detecting living objects in front of vehicles. The proposed and implemented prototype focuses on detecting squirrels and alerting the driver to avoid collision. This can be further extended to other small animals. The prototype uses transfer learning to leverage pre-existing knowledge from a larger, more robust model and apply it to a smaller, more specific task. The model is able to reach an accuracy of greater than 95% on color images with the help of this approach. The detection models are designed to be implemented on small, low-power devices integrated into vehicle electronics that are useful for object detection. The paper further focuses on a detailed comparison of different compression techniques to reduce the size of the data required for classification, enabling the system to run more efficiently and effectively. Further, the drop in accuracy and the total inference time are compared for different compression techniques. The study uniquely compares the different image compression and inference processing techniques which can be used as a base for future studies.