越線人群計數--Crossing-line Crowd Counting with Two-phase Deep Neural Networks
Crossing-line Crowd Counting with Two-phase Deep Neural Networks
ECCV2016
人群計數有兩種做法:1) region-of-interest (ROI) 感興趣區域計數,2) line-of-interest (LOI) 感興趣線計數,就是通過影象中某一條線的人流量。
大多數演算法關注 ROI 計數問題,但是我們認為 LOI 計數問題更有實際應用價值,對於人群密集的大型公共區域,我們想通過監控相機對所有區域進行人群計數是不現實的。但是我們可以對出口和入口進行 LOI 計數從而推理出公共區域裡的人數總量。
傳統的 LOI 方法是通過提取視訊中 對應 LOI 那條線上的資訊生成 2D temporal slices
generating 2D temporal slices by temporally concatenating video frame lines at the LOI
基於 temporal slices 我們來估計通過畫線的人流總量,這種方法在某些情況下效果不太好
When the scene is not crowded and pedestrians walks in normal speed (Fig. 1(b, row 1)), people can be well recognized in the temporal slices. However, we observe that temporal slices are not robust to scenes with high crowd density, slow walking speed, and low camera viewing angles. In Fig. 1(b, row2), the temporal slice shows excess jitters and people in it are no longer recognizable
本文使用 CNN 網路從視訊序列中學習特徵,分兩個步驟來解決計數問題
3 Method
3.1 Pixel-level supervision maps
根據視訊序列幀,我們得到中間結果 Density map 和 Velocity map,有了這兩個資訊,再去得到 Supervision crowd counting map
The crowd counting map can then be obtained as the elementwise multiplication of the density and velocity maps
相應位置數值相乘得到 crowd counting map
3.2 Deep Convolutional Neural Network for LOI crowd counting
我們的網路結構如下:
We propose to train the CNN in two phases. In the first training phase, the network is trained to predict the crowd density map D t and the crowd velocity map V t simultaneously
我們是分步訓練上面的網路的,首先訓練 predict the crowd density map 和 crowd velocity map, 再微調 fine-tune the trained network in the second phase with
supervision of the crossing-line crowd counting map
3.3 From crowd counting map to LOI counts
這一步主要是根據 crowd counting map 我們計算得到畫線的隨時間變化的人流量
4 Experiments
我們針對越線人群計數,建立了一個新的資料庫
1