Machine-Learning-Based Hazardous Spot Detection Framework by Mobile Sensing and Opportunistic Networks

This study proposes a framework to detect hazardous spots on roads by combining mobile sensing on commercial-use vehicles with vehicle-to-vehicle (V2V) opportunistic networking. The most likely hazardous spots are the locations where sudden braking or change in direction is necessitated. In the proposed system, a vehicle periodically makes a tentative decision regarding the state of the road using its own sensing data by a machine learning (ML)-based technique; however, the final decision is made by applying the Viterbi algorithm. A hidden Markov model (HMM) was introduced in the Viterbi algorithm for numerically representing the spatial relationship between the condition of the road and sequential output of the local hazardous spot detector. Numerical simulation analyses were conducted by using a real acceleration dataset of more than 25 commercial-use vehicles in Kakogawa City, Japan. It was shown that the Viterbi algorithm significantly improved the accuracy of detecting hazardous spots as compared to the tentative decisions made by support vector machines (SVMs). In addition, our study introduced a V2V cooperative data processing technique to achieve higher accuracy in the detection of perilous spots. The vehicles shared their local decisions through an opportunistic networking technology. Trellis pruning was integrated into the Viterbi algorithm based on the shared decisions for emitting the final decisions. It was demonstrated that the proposed V2V cooperative framework had the potential to enhance the accuracy of detecting hazardous spots for all the vehicles in a community.