Abstract
Modern technologies are being developed to address the alarming rise in road accidents caused by drivers’ errors. We leverage computer vision and deep learning at the edge (i.e., in a car) to detect vehicles and pedestrian that are in the surroundings. This information can then be employed to direct driver’s attention to relevant information, minimizing the effects of human errors. This work explores various deep learning pre-trained models: Intel open model zoo and TensorFlow detection model zoo to run inference on Intel Movidius to employ edge computing. We analyze the performance to determine the practicality of using the pre-trained model for road safety purposes. The experiments conducted examine the various Single Shot Multibox Detection (SSD) based models. The accuracy that we obtained by the weighted harmonic mean of the precision and recall on the models, the inference time and low demand in computing power determined that TensorFlow detection model zoo is a practical object detector that we can implement to tackle road safety issue.
Tagged: