Driver Assistance System for Detection of Marked Speed Breakers

Speed breaker detection is an important aspect of Driver assistance systems (DAS). Accidents caused by ignoring speed breakers can result in loss of life. This paper uses computer vision techniques and classical machine learning (ML) algorithms to detect marked speed breakers. A dataset of hand collected and web scraped images is compiled. These images undergo various pre-processing steps such as resizing and boundary estimation. An amalgamation of features is obtained from SIFT and HOG image feature descriptor algorithms which is further provided to the ML algorithms for training and classification. The final output is decided through a voting classifier system which incorporates classification outputs from the individual ML algorithms. This ensures that the classification is bias-free. A training accuracy of 97% and a test accuracy of 82% is obtained from the voting classifier. It also provides results with a precision score of 75%, recall score of 96%, and F1 score of 84%. The metrics showcase the capability of the system to detect marked speed breakers. The system is ported to Raspberry Pi 4 board with a camera-based input for real world implementation. A contemplation of these results is also carried out along with certain drawbacks that can be mitigated in further research endeavours.