Camera-based measurement of cyclist motion

Abstract

Heavy Goods Vehicles (HGVs) are overrepresented in cyclist fatality statistics in the UK relative to their proportion of total traffic volume. In particular, the statistics highlight a problem for vehicles turning left across the path of a cyclist on their inside. In this paper we present a camera-based system to detect and track cyclists in the blind spot. The system uses boosted classifiers and geometric constraints to detect cyclist wheels, and Canny edge detection to locate the ground contact point. The locations of these points are mapped into physical coordinates using a calibration system based on the ground plane. A Kalman Filter is used to track and predict the future motion of the cyclist. Full-scale tests were conducted using a construction vehicle fitted with two cameras, and the results compared with measurements from an ultrasonic-sensor system. Errors were comparable to the ultrasonic system, with average error standard deviation of 4.3 cm when the cyclist was 1.5 m from the HGV, and 7.1 cm at a distance of 1 m. When results were compared to manually extracted cyclist position data, errors were less than 4 cm at separations of 1.5 m and 1 m. Compared to the ultrasonic system, the camera system requires simpler hardware and can easily differentiate cyclists from stationary or moving background objects such as parked cars or roadside furniture. However, the cameras suffer from reduced robustness and accuracy at close range, and cannot operate in low-light conditions.

Multi-scale traffic sign detection model with attention

“…16 Zhang et al 17 used multiple scale filters in conjunction with edge detection lane lines while using adaptive threshold information to aid detection. Chris Eddy et al 18 used shape and edge detection techniques for bicycle pedestrian detection to reduce bicycle accidents. These feature-based methods have improved the detection performance of traffic signs to some extent.…”