Faroog Abdel-kareem Ibrahim

A Radar Guided Vision System for Vehicle Validation and Vehicle Motion Characterization

This paper describes a radar-guided monocular vision system that detects, validates, and tracks the preceding vehicle and thus predicts its lane-change intentions. A vision-based lane tracking process is developed to create a stable motion model in order to map the radar targets to image coordinates and consequently generate the region of interest (ROI) to search for a potential preceding vehicle. Model-based object classification algorithms are then applied to validate the existence of a vehicle in this ROI. Once the detected primary target vehicle, which is in the same lane as the host vehicle, is validated, it will be continuously tracked until it leaves the host lane. The spatial-temporal tracking history of the primary target vehicle is used to infer its intention of changing lanes. This prediction results from a characterization process for the primary target motion vector. The radar-vision integrated system has been evaluated on real-world data collected using a test vehicle equipped with a radar sensor, vision sensor, and a host processor.

Radar-based Target Tracking Method: Application to Real Road

Principle of the target tracking method for the Adaptive Cruise Control (ACC) system, which is applicable to nonuniform or transient condition, had been proposed by one of the authors. This method does not need any other information rather than that from the radar and host vehicle. Here the method is modified to meet more complex traffic scenarios and then applied to data measured on real highway. The modified method is based on the phase chart between the lateral component of the relative velocity and azimuth of a preceding vehicle. From the trace on the chart, the behavior of a preceding vehicle is judged and the discrimination between the lane change and curve-entry/exit can be made. The method can deal with the lane-change of a preceding vehicle on the curve as well as on the straight lane. And it is applied to more than 20 data including several road/vehicle conditions: road is straight, or turns right or left; vehicles are motorbikes, sedans and trucks. The algorithm could identify a target vehicle during the lane change and curve-entry/exit successfully. The trace of the data on the chart was found to be similar to what is expected in the analysis.

Optimal linear neuron learning and Kalman Filter based backpropagation neural network for DGPS/INS integration

This paper introduces an optimal least mean square (LMS) rule for a linear neuron DGPS/INS integration method. The optimal LMS rule is based on an online calculation of the learning rate based on the minimum variance criteria. Then, using this rule, the neuron adaptively estimates scale factor and the bias INS error source values to optimally combine the DGPS with INS. A similar concept of optimality is used to derive a Kalman filter based backpropagation training rule for a neural network DGPS/INS integration method. This method facilitates the use of the extended Kalman filter trained backpropagation neural network training method, which achieves an optimal training criterion. The mathematical derivations for both methods are introduced in this work. The performance of these methods for the INS error sources estimation is also demonstrated using real DGPS/INS data.