Novel classification of slow movement objects in urban traffic environments using wideband pulse Doppler radar

Every year thousands of people are involved in traffic accidents, some of which are fatal. An important percentage of these fatalities are caused by human error, which could be prevented by increasing the awareness of drivers and the autonomy of vehicles. Since driver assistance systems have the potential to positively impact tens of millions of people, the purpose of this research is to study the micro-Doppler characteristics of vulnerable urban traffic components, i.e. pedestrians and bicyclists, based on information obtained from radar backscatter, and to develop a classification technique that allows automatic target recognition with a vehicle integrated system. For this purpose, our methodology includes a detailed analysis of a commercial 76 GHz automotive radar system, an analysis of the state of the art on pedestrian classification with automotive radar, kinematic modeling of the targets of interest, simulations based on a theoretical model, empirical data analysis, classification features design and analysis, and implementation of a classification approach. Field data were collected in a controlled urban scenario where traffic was limited to the targets of interest. Following an initial comparison between the field data and simulated data micro-Doppler signatures, we observed that the commercial radar’s detection threshold prevented the clustering of low reflectivity target components. To be able to cluster these reflections, as they constitute a critical component on the success of our analysis, we artificially increased the reflectivity of some parts of the targets, i.e. pedestrian and bicyclist legs, with aluminum foil. The commercial radar system provides a number of target attributes including range and azimuth angle, from which range and cross range profiles are calculated, as well as amplitude, radial velocity, and related measurement statistics. Using our designed feature vectors - label data pairs, we train a neural network to perform automatic target discrimination. The hidden layer neurons perform a sigmoid function on its inputs, and the output layer neurons are unthresholded rectifier units. To maximize the utility of the labeled data and limit overfitting issues, we use a 10-fold cross validation technique to train and test the network. The resulting classifier system is capable of discriminating between three types of urban traffic components - cars, pedestrians, and bicyclists targets - based on fifteen target attributes with up to 91.1% accuracy. In this instance, a heuristic cost-classification method is exploited to maximize the protection 3 of vulnerable road users. Finally, given current limitations of the radar system, we offer a series of recommendations that would result in better classification capabilities for complex urban scenarios