




【摘要】為提升碰撞預(yù)警系統(tǒng)對周圍環(huán)境的感知能力,提出一種基于YOLOv5及危險區(qū)域判斷的碰撞預(yù)警系統(tǒng)。首先,通過通道注意力模塊提高模型的判別能力和準(zhǔn)確性,然后,使用路徑聚合網(wǎng)絡(luò)與空間金字塔池化提高模型對多尺度特征的提取能力,最后,通過引入預(yù)警激活區(qū)域過濾相對安全的目標(biāo),提高了預(yù)警系統(tǒng)的預(yù)警精確度。結(jié)果表明,引入預(yù)警激活區(qū)域后,與無預(yù)警激活區(qū)域相比,預(yù)警系統(tǒng)的準(zhǔn)確度、精度和召回率分別提高20%、50%和26.7%,運行速度提升49.1%,進一步證明了方法的有效性。
主題詞:YOLOv5 通道注意力模塊 路徑聚合網(wǎng)絡(luò) 空間金字塔池化 預(yù)警激活區(qū)域
碰撞預(yù)警系統(tǒng)
中圖分類號:U463.6;TP291.3" 文獻標(biāo)志碼:A" "DOI: 10.19620/j.cnki.1000-3703.20240026
Research on Collision Warning System Based on Improved YOLOv5 amp; Hazardous Zone Judgment
Yi Zhenxing1, Zhan Zhenfei1,2, Mao Qing1, Sun Bowen1, Wang Ju2
(1. Chongqing Jiaotong University, Chongqing 400074; 2. State Key Laboratory of Automotive Noise, Vibration and Safety Technology, Chongqing 401120)
【Abstract】In order to improve the ability of the collision warning system to perceive the surrounding environment, this paper proposed a collision warning system based on YOLOv5 and hazardous area judgment. Firstly, the discriminative ability and accuracy of the model were improved by the channel attention module, then, the extraction ability of the model for multi-size features was improved by using path aggregation network and spatial pyramid pooling, and finally, the warning accuracy of the warning system was improved by filtering relatively safe targets through the introduction of warning activation regions. The results show that the introduction of warning activation regions improves the accuracy, precision and recall of the warning system by 20%, 50% and 26.7%, respectively, the running speed is increased by 49.1%, which further proves the effectiveness of the method.
Key words: YOLOv5, Channel attention module, Path aggregation network, Spatial pyramid pooling, Warning activation area, Collision warning system
1 前言
車輛碰撞預(yù)警系統(tǒng)通過實時監(jiān)測分析車輛周圍的環(huán)境信息,及時發(fā)出警示,從而避免碰撞事故發(fā)生。
近年來,隨著計算機視覺、傳感器技術(shù)和機器學(xué)習(xí)的快速發(fā)展,基于目標(biāo)檢測的碰撞預(yù)警系統(tǒng)一般通過改進目標(biāo)檢測方法與改進預(yù)警策略提高系統(tǒng)性能。在改進目標(biāo)檢測方法方面:談文蓉[1]通過改進YOLOv4算法提高碰撞預(yù)警系統(tǒng)的準(zhǔn)確性和實時性,但是無法檢測出其他交通參與者,如行人、非機動車等;夏聰[2]采用基于單目視覺的車輛前向碰撞預(yù)警系統(tǒng),實現(xiàn)了車輛前向行人與非機動車碰撞預(yù)警,但在檢測小目標(biāo)或密集目標(biāo)時,該方法檢測性能較差;胡超超[3]基于YOLOv2網(wǎng)絡(luò)引入殘差網(wǎng)絡(luò),提高了模型檢測群簇小目標(biāo)的準(zhǔn)確率,但目標(biāo)與背景相近或被嚴(yán)重遮擋時,極易出現(xiàn)漏檢現(xiàn)象?!?br>