基于行车声音端点检测的交通量统计

基于传统特征的行车声音端点检测法存在重叠有车段识别率低、双门限阈值较难确定的问题,针对这两个问题,探索性地将梅尔频率倒谱系数(Mel frequency cepstral coefficients,MFCC)倒谱距离特征和短时能量特征进行了融合并应用于交通量检测。首先选取了周围环境较为安静的一个双车道路段,并采集了该路段上包含重叠有车段的行车声音;其次提取了行车声音的短时能量特征和MFCC倒谱距离特征,并对它们在端点检测中的优劣进行了分析对比;再次提出了一种融合短时能量特征和MFCC倒谱距离特征的新特征,并基于新特征将传统的双门限判决思路改进成了单门限判决思路;最后利用新特征对有车段进行端点检测并统计交通量。实验结果表明:基于融合特征的端点检测方法能有效解决重叠有车段识别率低和双门限阈值较难确定的问题。

Traffic Statistics Based on the Endpoint Detection of Driving Acoustic Signals

Based on traditional characteristics, the endpoint detection method of driving acoustic signals has some problems, such as low recognition rate of overlapping vehicle segment and difficulty in determining double threshold value. For the two problems, mel frequency cepstral coefficients (MFCC) cepstrum distance and short-term energy were integrated to detect traffic. First, a dual-lane road with relatively quiet environment was selected, and the driving acoustic signals containing overlapping vehicle segments were collected from the road. Second, the short-term energy and MFCC cepstrum distance were extracted. Their advantages and disadvantages of endpoint detection were analyzed and compared. Third, a new feature which integrates the short-term energy and MFCC cepstrum distance was proposed. Based on the new feature, the traditional dual-threshold decision was improved to single-threshold decision. Finally, the new feature was used to detect the vehicle segment endpoints and to count the traffic volume. The experimental results show that the endpoint detection method based on integrated feature can effectively solve the problems of low recognition rate and difficult determination of double threshold.