Vision-based Multiple Moving Objects Detection for Intelligent Automobiles

This paper presents a vision-based multiple moving objects detection work which attracts much attention in intelligent automobile applications recently. This system provides behavior information of objects and is an intuitive detection method similar to human visual perception. Besides, vision-based objects detection methods are much low-cost compared with other detection methods such as RADAR (Radio Detection and Ranging), or LiDAR (Light Detection and Ranging). However, current vision-based objects detection methods still suffer from several challenges such as high false alarm rate and unstable detection rate which limit their value in practical applications. A well-known machine learning algorithm, i.e. AdaBoost, has been adopted in this detection system but the original system reveals the aforementioned problems, as well. After thorough exploration and observation, we found that the way we train the classifier affects the detection result obviously. Accordingly, this paper suggests a set of sample selection rules and a multi-pass self-correction training process for effective vision-based moving objects detection except a detailed literature survey of recent state-of-the-arts. According to the best of our knowledge, the parts of sample selection rules and self-correction training process distinguish this paper from the others in the literature. In addition, a prototype of the presented detection system is also realized on a portable platform equipped with a 1 GHz ARM Cortex-A9 RISC processor only. The experimental results show that the presented detection system can successfully detect multiple moving objects including pedestrians, motorcycles, and vehicles commonly appeared on the road with 91.8% detection rate and 3.3% false alarm rate.

並列關鍵字

Intelligent automobile ； intelligent vision ； object detection ； pedestrian detection

國際替代計量

全文下載

主題瀏覽