Intelligent video surveillance: a review through deep learning techniques for crowd analysis

The review include methods which are having deep learning background and methods which are not having that background.

Spatial temporal convolutional neural networks for anomaly detection and localization in crowded scenes [114] shows the problem related with crowd analysis is challenging because of the following reasons

• Large number of pedestrians

• Close proximity

• Volatility of individual appearance

• Frequent partial occlusions

• Irregular motion pattern in crowd

• Dangerous activities like crowd panic

• Frame level and pixel level detection

The paper suggests optical flow based solution. The CNN is having eight layers. Training is based on BVLC caffe. Random initialization of parameters is done and system is trained through stochastic gradient descent based back propagation. The implementation part is done by considering four different datasets like UCSD, UMN, Subway and finally U-turn. The details of implementation regarding UCSD includes frame level and pixel level criterion. Frame level criterion concentrates on temporal domain and pixel level criterion considers both spatiial and temporal domain. Different metrics to evaluate performance includes EER (Equal Error Rate) and Detection Rate (DR).

Online real time crowd behavior detection in video sequences [115] suggests FSCB, behavior detection through feature tracking and image segmentation. The procedure involves following steps

• Feature detection and temporal filtering

• Image segmentation and blob extraction

• Activity detection

• Activity map

• Activity analysis

• Alarm

The main advantage is no need of training stage for this method. The method is quantitatively analyzed through ROC curve generation. The computational speed is evaluated through frame rate. The data set considered for experiments include UMN, PETS2009, AGORASET and Rome Marathon.

Deep learning for scene independent crowd analysis [82] proposes a scene independent method which include following procedures

• Crowd segmentation and detection

• Crowd tracking

• Crowd counting

• Pedestrian travelling time estimation

• Crowd attribute recognition

• Crowd behavior analysis

• Abnormality detection in a crowd

Attribute recognition is done thorugh a slicing CNN. By using a 2D CNN model learn appearance features then represent it as a cuboid. In the cuboid three temporal filters are identified. Then a classifier is applied on concatenated feature vector extracted from cuboid. Crowd counting and crowd density estimation is treated as a regression problem. Crowd attribute recognition is applied on WWW Crowd dataset. Evaluation metrics used are AUC and AP.

The analysis of High Density Crowds in videos [80] describes methods like data driven crowd analysis and density aware tracking. Data driven analysis learn crowd motion patterns from large collection of crowd videos through an off line manner. Learned pattern can be applied or transferred in applications. The solution includes a two step procedure. Global crowded scene matching and local crowd patch matching. Figure 2 illustrates the two step procedure.

The database selected for experimental evaluation includes 520 unique videos with 720 × 480 resolutions. The main evaluation is to track unusual and unexpected actions of individuals in a crowd. Through experiments it is proven that data driven tracking is better than batch mode tracking. Density based person detection and tracking include steps like baseline detector, geometric filtering and tracking using density aware detector.

A review on classifying abnormal behavior in crowd scene [77] mainly demonstrates four key approaches such as Hidden Markov Model (HMM), GMM, optical flow and STT. GMM itself is enhanced with different techniques to capture abnormal behaviours. The enhanced versions of GMM are

• GMM

• GMM and Markov random field

• Gaussian poisson mixture model and

• GMM and support vector machine

GMM architecture includes components like local descriptor, global descriptor, classifiers and finally a fusion strategy. The distinction between normal and and abnormal behaviour is evaluated based on Mahalanobis distance method. GMM–MRF model mainly divided into two sections where first section identifies motion pttern through GMM and crowd context modelling is done through MRF. GPMM adds one extra feture such as count of occurrence of observed behaviour. Also EM is used for training at later stage of GPMM. GMM–SVM incorporate features such as crowd collectiveness, crowd density, crowd conflict etc. for abnormality detection.

HMM has also variants like

• GM-HMM

• SLT-HMM

• MOHMM

• HM and OSVMs

Hidden Markov Model is a density aware detection method used to detect motion based abnormality. The method generates foreground mask and perspective mask through ORB detector. GM-HMM involves four major steps. First step GMBM is used for identifying foreground pixels and further lead to development of blobs generation. In second stage PCA–HOG and motion HOG are used for feature extraction. The third stage applies k means clustering to separately cluster features generated through PCA–HOG and motion–HOG. In final stage HMM processes continuous information of moving target through the application of GM. In SLT-HMM short local trajectories are used along with HMM to achieve better localization of moving objects. MOHMM uses KLT in first phase to generate trajectories and clustering is applied on them. Second phase uses MOHMM to represent the trajectories to define usual and unusual frames. OSVM uses kernel functions to solve the nonlinearity problem by mapping high dimensional features in to a linear space by using kernel function.

In optical flow based method the enhancements made are categorized into following techniques such as HOFH, HOFME, HMOFP and MOFE.

In HOFH video frames are divided into several same size patches. Then optical flows are extracted. It is divided into eight directions. Then expectation and variance features are used to calculate optical flow between frames. HOFME descriptor is used at the final stage of abnormal behaviour detection. As the first step frame difference is calculated then extraction of optical flow pattern and finally spatio temporal description using HOFME is completed. HMOFP Extract optical flow from each frame and divided into patches. The optical flows are segmented into number of bins. Maximum amplitude flows are concatenated to form global HMOFP. MOFE method convert frames into blobs and optical flow in all the blobs are extracted. These optical flow are then clustered into different groups. In STT, crowd tracking and abnormal behaviour detection is done through combing spatial and temporal dimensions of features.

Crowd behaviour analysis from fixed and moving cameras [78] covers topics like microscopic and macroscopic crowd modeling, crowd behavior and crowd density analysis and datasets for crowd behavior analysis. Large crowds are handled through macroscopic approaches. Here agents are handled as a whole. In microscopic approaches agents are handled individually. Motion information to represent crowd can be collected through fixed and moving cameras. CNN based methods like end-to-end deep CNN, Hydra-CNN architecture, switching CNN, cascade CNN architecture, 3D CNN and spatio temporal CNN are discussed for crowd behaviour analysis. Different datasets useful specifically for crowd behaviour analysis are also described in the chapter. The metrics used are MOTA (multiple person tracker accuracy) and MOTP (multiple person tracker precision). These metrics consider multi target scenarios usually present in crowd scenes. The dataset used for experimental evaluation consists of UCSD, Violent-flows, CUHK, UCF50, Rodriguez’s, The mall and finally the worldExpo’s dataset.

Zero-shot crowd behavior recognition [79] suggests recognizers with no or little training data. The basic idea behind the approach is attribute-context cooccurrence. Prediction of behavioural attribute is done based on their relationship with known attributes. The method encompass different steps like probabilistic zero shot prediction. The method calculates the conditional probability of known to original appropriate attribute relation. The second step includes learning attribute relatedness from Text Corpora and Context learning from visual co-occurrence. Figure 3 shows the illustration of results.

Computer vision based crowd disaster avoidance system: a survey [81] covers different perspectives of crowd scene analysis such as number of cameras employed and target of interest. Along with that crowd behavior analysis, people count, crowd density estimation, person re identification, crowd evacuation, and forensic analysis on crowd disaster and computations on crowd analysis. A brief summary about benchmarked datasets are also given.

Fast Face Detection in Violent Video Scenes [83] suggests an architecture with three steps such as violent scene detector, a normalization algorithm and finally a face detector. ViF descriptor along with Horn–Schunck is used for violent scene detection, used as optical flow algorithm. Normalization procedure includes gamma intensity correction, difference Gauss, Local Histogram Coincidence and Local Normal Distribution. Face detection involve mainly two stages. First stage is segmenting regions of skin and the second stage check each component of face.

Rejecting Motion Outliers for Efficient Crowd Anomaly Detection [54] provides a solution which consists of two phases. Feature extraction and anomaly classification. Feature extraction is based on flow. Different steps involved in the pipeline are input video is divided into frames, frames are divided into super pixels, extracting histogram for each super pixel, aggregating histograms spatially and finally concatenation of combined histograms from consecutive frames for taking out final feature. Anomaly can be detected through existing classification algorithms. The implementation is done through UCSD dataset. Two subsets with resolution 158 × 238 and 240 × 360 are present. The normal behavior was used to train k means and KUGDA. The normal and abnormal behavior is used to train linear SVM. The hardware part includes Artix 7 xc7a200t FPGA from Xilinx, Xilinx IST and XPower Analyzer.

Deep Metric Learning for Crowdedness Regression [84] includes deep network model where learning of features and distance measurements are done concurrently. Metric learning is used to study a fine distance measurement. The proposed model is implemented through Tensorflow package. Rectified linear unit is used as an activation function. The training method applied is gradient descent. Performance is evaluated through mean squared error and mean absolute error. The WorldExpo dataset and the Shanghai Tech dataset are used for experimental evaluation.

A Deep Spatiotemporal Perspective for Understanding Crowd Behavior [61] is a combination of convolution layer and long short-term memory. Spatial informations are captured through convolution layer and temporal motion dynamics are confined through LSTM. The method forecasts the pedestrian path, estimate the destination and finally categorize the behavior of individuals according to motion pattern. Path forecasting technique includes two stacked ConvLSTM layers by 128 hidden states. Kernel of ConvLSTM size is 3 × 3, with a stride of 1 and zeropadding. Model takes up a single convolution layer with a 1 × 1 kernel size. Crowd behavior classification is achieved through a combination of three layers namely an average spatial pooling layer, a fully connected layer and a softmax layer.

Crowded Scene Understanding by Deeply Learned Volumetric Slices [85] suggests a deep model and different fusion approaches. The architecture involves convolution layers, global sum pooling layer and fully connected layers. Slice fusion and weight sharing schemes are required by the architecture. A new multitask learning deep model is projected to equally study motion features and appearance features and successfully join them. A new concept of crowd motion channels are designed as input to the model. The motion channel analyzes the temporal progress of contents in crowd videos. The motion channels are stirred by temporal slices that clearly demonstrate the temporal growth of contents in crowd videos. In addition, we also conduct wide-ranging evaluations by multiple deep structures with various data fusion and weights sharing schemes to find out temporal features. The network is configured with convlutional layer, pooling layer and fully connected layer with activation functions such as rectified linear unit and sigmoid function. Three different kinds of slice fusion techniques are applied to measure the efficiency of proposed input channels.

Crowd Scene Understanding from Video A survey [86] mainly deals with crowd counting. Different approaches for crowd counting are categorized into six. Pixel level analysis, texture level analysis, object level analysis, line counting, density mapping and joint detection and counting. Edge features are analyzed through pixel level analysis. Image patches are analysed through texture level analysis. Object level analysis is more accurate compared to pixel and texture analysis. The method identifies individual subjects in a scene. Line counting is used to take the count of people crossed a particular line.

Table 10 will discuss some more crowd analysis methods.