Novel Approach in Labeling Data for Classification of Warning Level While Driving

A novel approach for predicting collision warning levels using a classifier that interprets a window of present and past relative positions between the ego vehicle and front object is considered. The premise assumes that the ego vehicle is equipped with a LiDAR, radar or camera that measures the relative distance. A machine learning approach is presented where the feature is the time-to-collision (TTC) which can be determined from Kalman filter estimate of distance, speed and acceleration. The classifier produces classes of warning labels that combines alert levels with watch levels. First part of the classifier called predictor classifier 1, produces the alert levels by evaluating the TTC value of current instance; however, the second part of the classifier called predictor classifier 2, produces the watch labels by applying a novel mathematical algorithm that interprets past values of TTC in a given window. Alert and watch labels are merged together by warning labels generator to produce the final labels for the dataset. This paper presents the formulation and simulation results for the machine learning approach that utilizes neural networks. A future extension of the result will address a deep learning approach where the Kalman filtering & TTC features will be eliminated.