An intelligent campus simulation system based on game

The dataset used in this paper is obtained from the CASIA behaviour analysis dataset of the Chinese Academy of Sciences, which consists of 422 video data with a frame rate of 25 fps. The dataset is shot from three angles, that is, horizontal angle, oblique angle and overlooking angle with single-person behaviour and multi-person interaction behaviour. Each video is recorded between 5 s and 30 s, depending on different behaviours.

There are other behaviours unrelated to this behaviour in the video. To avoid affecting the model performance by the irrelevant behaviour, it is initially needed to delete these behaviours in the video. Since there are too many overlaps between adjacent frames, there is a lot of overlapping information between them. Thus, a lot of redundant calculations will be encountered when extracting features, which affects the performance of the model. To reduce redundant information and improve operational efficiency, this paper employs the method of extracting N-frame images at the same time interval.

The process of training a model is very complex and requires a lot of calculations. If all kinds of behavioural video data are insufficient, it is easy to cause model over-fitting. In general, the larger the data volume, the more robust the model is. To increase the generalisation ability of the model, we use the method of data augmentation to increase the data amount Before training the model, the data is augmented. We use six augmentation methods to augment and save the augmented results, that is, the upper right corner cutting, upper left corner cutting, lower right corner cutting, lower left corner cutting, centre cutting, and horizontal flip of the above enhancement results and the original image, the final data set is enlarged 11 times than the original data set size, which can increase the robustness of the model, and provide enough data set for training. In each training round, the pictures are randomly cut to 112×112 and fed into the network.

The 3DCNN in this paper are made up of overlapping convolutional and pooling layers. The 3DCNN, different from 2DCNN, can not only extract spatial features in the video but also learn temporal information in it. The pooling layer downsamples, removes redundant features and presses features to reduce computation. The network takes the 16-frame pictures of 16×112×112×3 as the input and obtains the feature vectors through eight convolution layers and five pooling layers. Whereinto, the convolution core size is 3×3×3 and the step size is 1. In the pooling layer, all the pooling window is 2×2×2 with 2×2×2 step size except the first pool layer which is 1×2×2 in pooling window and 1×2×2 in step size.

The CBAM [8] attention mechanism model contains two independent sub-modules, namely, the channel attention module (CAM) and the spatial attention module (SAM). It is a lightweight attention module, with negligible cost when added to other models, and can be trained in an end-to-end manner together with the original model. Woo et al. [8] added it to ResNet and MobileNet, significantly improving the model performance. For a given feature map, CBAM will compute the attention map along the path attention channel and the spatial attention channel, and then multiply the attention map by the input feature map to conduct feature learning. In order that the 3DCNN can learn features better, the CBAM model can be improved to be combined with 3DCNN. The specific refined model is shown in Figure 2.

Fig. 2

After entering the input feature F_3D ∈ R^T*W*H*C, and passing the path attention module M_c3D ∈ R^1*1*1*C, the convolution result multiplies the input feature to obtain the output of the path attention module, which is input to the SAM M_s3D ∈ R^T*W*H*1. By multiplying the result with the input feature, the final output result of the improved attention model is obtained. The formula is as follows: (1) F3D1=Mc3D(F3D)⊗F3D {\rm{F}}_{3{\rm{D}}}^1 = {{\rm{M}}_{{\rm{c}}3{\rm{D}}}}\left( {{{\rm{F}}_{3{\rm{D}}}}} \right) \otimes {{\rm{F}}_{3{\rm{D}}}} (2) F3D2=Ms3D(F3D1)⊗F3D1 {\rm{F}}_{3{\rm{D}}}^2 = {{\rm{M}}_{{\rm{s}}3{\rm{D}}}}\left( {{\rm{F}}_{3{\rm{D}}}^1} \right) \otimes {\rm{F}}_{3{\rm{D}}}^1

CAM mainly focuses on which paths contribute to the final classification. It adopts global average pooling and global maximum pooling. The two pooling are more conducive to extracting richer high-level features. The specific formula is as follows: (3) Mc3D(F3D)=σ(MLP(AvgPoll3D(F3D))+MLP(MaxPoll3D(F3D))) =σ(W1(W0(Favgc))+W1(W0(FMaxc))) \matrix{ {{{\rm{M}}_{{\rm{c}}3{\rm{D}}}}\left( {{{\rm{F}}_{3{\rm{D}}}}} \right) = \sigma \left( {{\rm{MLP}}\left( {AvgPoll3D\left( {{F_{3D}}} \right)} \right) + {\rm{MLP}}\left( {MaxPoll3D\left( {{F_{3D}}} \right)} \right)} \right)} \hfill \cr {\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\; = \sigma \left( {{W_1}\left( {{W_0}\left( {F_{avg}^c} \right)} \right) + {W_1}\left( {{W_0}\left( {F_{Max}^c} \right)} \right)} \right)} \hfill \cr }

First, the feature F_3D is entered, and two features of 1 × 1×1×C are obtained by the two pooling, and sent to the MLP network. After summing the corresponding elements of the output of MLP and passing the sigmoid activation function, the final feature is obtained, which is multiplied with the input feature F_3D to obtain the feature required by the CAM.

SAM mainly focuses on which positions in the image can affect the final classification. It also adopts global average pooling and global maximum pooling. The specific formula is as follows: (4) Ms3D=σ(f7*7([AvgPoll3D(F3D1);MaxPoll3D(F3D1)]))=σ(f7*7([fAvgs;fMaxs])) {{\rm{M}}_{{\rm{s}}3{\rm{D}}}} = \sigma \left( {{{\rm{f}}^{7*7}}\left( {\left[ {AvgPoll3D\left( {F_{3D}^1} \right);MaxPoll3D\left( {F_{3D}^1} \right)} \right]} \right)} \right) = \sigma \left( {{{\rm{f}}^{7*7}}\left( {\left[ {f_{Avg}^s;f_{Max}^s} \right]} \right)} \right)

Whereinto F3D1 {\rm{F}}_{3{\rm{D}}}^1 is the output feature of the CAM used as the input to the SAM. After passing through two pooling layers, F3D1 {\rm{F}}_{3{\rm{D}}}^1 obtains two T×H×W×1 features, and then the two features are stitched together. After a 7×7 convolution, then the final feature is obtained through sigmoid, and this feature and the input feature F3D1 {\rm{F}}_{3{\rm{D}}}^1 are multiplied to obtain the output feature.

Define 1 multi-agent mobile target tracking strategy selected model Multi-Agent Deep Deterministic Policy Gradient (MADDPG), described by six-agent group (E, A, D, P, P˜ {\rm{\tilde P}} , U).

The runaway action is first assigned a prior probability that can be obtained by historical experience or equally assigned probability [10]. The manchaser adjusts his tracking strategy by observing the escape behaviour and constantly correcting the judgement of the runaway's position. The game order is as follows:

Select an action Ei from the action space Ea of the runaway Na at a certain probability, where Ei∈ Ea. The runaway Na knows Ei and manchaser Nd does not, but the manchaser Nd has the inference on Ei, that is to say, the manchaser knows the prior probability of the runaway's action.

The calculation steps of the refine Bayes’ equilibrium solution are as follows:

The manchaser establishes posterior probability inference p˜(θi|aj) {\rm{\tilde p}}\left( {\theta {\rm{i}}|{\rm{aj}}} \right) on each meshed location information set.

The subgame refined Nash equilibrium (a * (θ), d * (a)) gained in steps (2) and (3) is used to obtain the inference p˜*=p˜*(θ|a) {{\rm{\tilde p}}^*} = {{\rm{\tilde p}}^*}(\theta |{\rm{a}}) of the location information of the runaway satisfying the Bayes’ rule. If p˜*(θ|a) {{\rm{\tilde p}}^*}(\theta |{\rm{a}}) does not conflict with P˜*(θ|a) {{\rm{\tilde P}}^*}(\theta |{\rm{a}}) , the refined Bayes’ equilibrium is solved as: (7) EQ=(a*(θ),d*(a),p˜*(θ|a)) EQ = \left( {a*\left( \theta \right),d*\left( a \right),{\rm{\tilde p}}*\left( {\theta |a} \right)} \right)

According to the game theory [13], the strategy in equilibrium is the optimal choice for both sides, so the manchaser believes that d * (a) is chosen as the optimal hunting strategy.

In the experimental environment, n manchasers must cooperate with each other to chase a runaway in a 2-dimensional discrete k×k grid environment. We assume that the grid is bounded by the wall, that is, if moving in a certain direction causes the manchaser or the runaway to hit the wall, the movement fails, and the agent remains in the same grid. The environmental state includes the positions of all manchasers d1t, d2t,…, dnt and the location of the runaway gt. This ambient diversely runs, and for each episode, the manchaser's initial locations d10, d20,…, dn0 and the runaway's initial position g0 are uniformly sampled, so that no two locations are the same. Each round is composed of many time steps. At each time step, the manchaser performs one of the five possible actions: up, down, left, right, or static. If there is more than one manchaser moving into the same grid, the movement fails and the manchasers remain in their own grids. For models of the reward function, we consider the following settings of the reward function:

Sparse team reward: At each time step, all manchasers are punished by 1 for delaying capturing the runaway, and zero punishment if successful.

The combinatorial space of the hyperparameters is too large for an exhaustive search. Due to high computational costs, we do not perform a systematic grid search. Instead, we only conduct informal searches. We linearly anneal from 1.0 to 0.01 during the exploration phase (100,000 time step) and fix it at 0.01 thereafter. For the first environment, we train each agent's network for a total of 50,000 sets, while for the second environment, each network is trained for 20,000 sets. We use replay memory with 10,000 recent time steps. The RMSProp algorithm that we use has a learning rate of 0.00005. Furthermore, we use a discount factor of 0.99 and perform a learning update every 4 time steps using a small batch of 32 transformations. For replay priority, we use the priority index ω of 0.5 and linearly increase the importance sampling (IS) index from 0.4 to 1 during the exploration phase (100,000 time steps). The specific parameters of the experiment are shown in the following figure:

During training, the manchaser and the runaway move for five movements like up and down, left and right, or static. The algorithm is trained 500 times, and the cumulative return of each time is recorded. The results are shown in the Figure. It can be seen from Figure 4 that the algorithm starts to converge when trained about 100 times.

Fig. 4

Through analysing the attack and defence game model and analysing the proceeds of the algorithm, the general law of the hunting strategy of the algorithm can be obtained: the accuracy of the detection of the escape strategy has a significant impact on the game trend of the hunting and the equilibrium solution. When the manchaser may mistake the escape strategy, introducing the minimum risk strategy benefit is closer to the actual hunting situation so that the algorithm can choose a better hunting strategy. Therefore, the proposed algorithmic strategy should be deployed together with the hunting detection strategy to achieve optimal hunting effects.

In the research process of various safety issues, the campus, as one of the types with a crowded population, has a typical risk amplification and social response characteristics, making it have a very important safety significance. In this paper taking Xinjiang Normal University as an example, the campus map is shown in Figure 5a). MATLAB software is used to abstract important sites as nodes and numbers them, and abstract the roads between important sites as edges. The obtained campus network map is shown in Figure 5b), which consists of 44 nodes and 64 edges.

Fig. 5

Based on the experimental content in Section 4.2, the planning platform is developed to set the initial position of attacker escape g0 (point 39 as shown in Figure 5b) as the attack point, requiring a certain amount of security resources, and the material storage in the real campus background is the starting point of path planning, points 2, 22 and 31 as shown in Figure 5b). The simulation results are shown in Figure 6 (bold blue line is the acquired optimal path):

Fig. 6