Vehicle Accident Detection using alternating Convolutional and Max-pooling layers in a CNN

Abstract: According to the WHO, approximately 1.3 million people die every year due to traffic related accidents. One of the factors determining survival probability is the response time of the paramedics. This variable depends on the actions of onsite pedestrians or the person responsible for traffic proceedings at the site of accident. We propose the use of a CNN with alternating max-pooling layers to classify images as accident and non-accident to automatically detect an accident over CCTV camera footage. One problem plaguing this field is the lack of available accident datasets, so we prepare the dataset from various sources. The results (training accuracy of 93.78% and validation accuracy of 82.65%) show that the use of alternating max-pooling layers help in accident detection even when the vehicle is not in the centre of the frame.

Keywords: accident detection, convolutional neural network, image classification, deep learning