Abstract:
The number of deaths from car accidents has gone up a lot lately. This trend is likely to continue, especially with more people and cars on the road. Delays in reporting accidents can lead to serious injuries if emergency vehicles and response teams are slow to reach the crash site. Various IoT systems have been developed in order to present a solution to the problem of delayed access to medical care due to the time it takes to alert response teams. One of the most important objectives in developing these solutions is to make sure that accidents are detected accurately to avoid false alerts and alarms while coming up with a solution. This research thesis focuses on the development of a real time road accident detection system using tiny machine learning (TinyML). This technique brings the power of machine learning to resource-constrained devices, enabling realtime, low-latency inference at the edge with minimal power consumption and reduced data transmission needs. Therefore, this research thesis utilized TinyML in order to develop a system that detects road accidents in real time. An image dataset was collected for three classes; Accident Detected, No_Accident_Detected and Other Objects. Then we used MobileNetV1, a convolutional neural network (CNN) architecture with specific modifications. Input images were sized at 96x96 pixels, balancing efficiency and detail. A width multiplier of 0.2 reduced computational complexity for resource-constrained deployment, trading some accuracy. The final fully connected layer was removed as classification was unnecessary; only high-level feature extraction was required. A 10% dropout rate after the final convolutional layer regularized the model during training to prevent overfitting. This MobileNetV1 96x96 0.2 (no final dense layer, 0.1 dropout) configuration tailored the architecture for the research goals, emphasizing efficiency while removing extraneous components and incorporating regularization. The model was trained using the image dataset and we got an accuracy of 87.7% and an accuracy of 86.12% on new, unseen data. Finally, the model was deployed into Arduino Nano 33 BLE sense with on device performance of 431 milliseconds inferencing time, Peak RAM usage of 97.4KB and 220.9KB Flash usage. Future work will include an integration of a communication model to alert relevant authoritie
