Efficient Vehicle Accident Detection System using Tensorflow and Transfer Learning

In today’s era, need of efficient accident detection has drawn much attention as number of accidents are increasing day by day. One of the widely employed method is to use accelerometer to detect a crash. In this method, acceleration (g) value measured from the accelerometer is calibrated to detect an accident. This method, however is limited by the accuracy of the accelerometer. To make an efficient accident detection system, convolutional neural network (CNN) methodology can be incorporated in the system. CNN is the state-of-the-art method for image classification. In the recent work, image classification has been used to detect accident. However, CNN takes large time, data and computing power to be trained. To mitigate these issues, transfer learning technique has been innovatively incorporated for the accident detection application, which involves retraining the already trained network. Inception-v3 is an image classifier developed by google, which is incorporated for this purpose. In this work, accident detection system is designed using advanced and efficient Transfer Learning algorithm, which gives 84.5% of accuracy. Also, an effective comparison between this advanced method and the traditional accelerometer based technique have been made.