A double-region learning algorithm for counting the number of pedestrians in subway surveillance videos

Dynamic Region Division for Adaptive Learning Pedestrian Counting

“…Figure 2 clearly shows the flow chart of algorithm. Since we find that gaussian kernels are not suitable for simulating large heads, basing on the straight-line double region pedestrian counting method [7], we propose a dynamic region division algorithm to keep the completeness of counting objects. Utilizing the object bounding boxes obtained by YoloV3 and expectation division line of the scene, the boundary for nearby region and distant one is generated under the premise of retaining whole head.…”

“…To maximize advantages of different methods in corresponding regions, it is significant to divide regions properly. He et al [7] used a straight line to divide, but it causes the error that one head may be cut into two parts by the line. To avoid this problem, we propose a dynamic region division algorithm.…”

“…We evaluate the proposed algorithm through extensive experiments on the publicly available Subway station pedestrian dataset [7]. The dataset covers five typical subway station scenes in Shanghai.…”

“…To augment the training set for training the ID-CNN, we flip the frames horizontally to double the training set. Following [7], we use the absolute error(M AE) as the evaluation metric which are defined as follows:…”

“…Therefore, using density-based method in high perspective distortion scenes can not obtain a optimal counting result. All the aforementioned methods consider only one method for entire frame, while He et al [7] proposed a straight-line double region pedestrian counting method. It make the best use of features from different regions to design appropriate counting method.…”