Walking path detection of pedestrians by using hard negative learning

Parallel to the developing technology, the use of autonomous vehicles is increasing rapidly. An autonomous vehicle has to be able to move safely within city traffic and there are some situations that need to be addressed. An autonomous vehicle flowing in traffic must be aware of its own situation, behave safe and effective like a driver, and be able to make critical decisions. Autonomous vehicle must use the past experience as well as the current state to predict future states. In this study, Histogram of Directional Gradients (HOG) algorithm and Support Vector Machine (SVM) are used to predict pedestrians' routes to make the autonomous cars safer. It is aimed to make their driving safer. A camera located inside the vehicle to capture the video signal. For each frame subtracted from the resulting video images, positions of the pedestrians are calculated and the camera's self-movement is eliminated. After the self-movement is eliminated, the routes of the pedestrians are predicted by using frame by frame position changes. With the help of the average of the movement found, the future positions of the pedestrians are predicted. The proposed method was applied on Daimler Mono Pedestrian Benchmark dataset. The results are promising regarding to the literature. If Gaussian Kernel SVM is used, 95.27 per cent of the pedestrians were detected. 96.69 per cent of the routes of detected pedestrians were correctly predicted.