Study On Fire Detection And Body Shape Recognition Based On Depth Images

In this paper, the study based on depth image technology of its application on fire detection and body recognition were studied with TOF (time of flight) camera acted as the main facility. Fire detection algorithm and body recognition algorithm were developed, and their corresponding recognition rates were with higher accuracy. A fire supervisory system based on depth image technology was established by integrating the algorithm and facilities that put the study results into practice. The main contents of this paper are listing as follows:In order to develop the application of time of flight algorithm in fire detection and simplify the algorithm to improve detection rate and accuracy, according to the time of flight depth map method, considering the characteristics of depth map of fire flame, fire flame identification algorithm based on variation rate of time of flight depth map was designed. Several groups of fire flame identification experiments, including heptane flame, ethanol flame, paper flame, lamplight interference and pedestrian interference test, were carried out with3-D depth camera acted as main equipment. The captured maps were processed and computed. A simplified algorithm was designed for fire flame identification which was used to analyze the characteristics of depth map, frequency spectrogram, concentration ratio and area fluctuation of fire flame. The results indicate that the identification precision rate is greater than91.5%, and the misrecognition rate is less than3.8%. Fire flame could be efficiently identified with this algorithm.Secondly, body recognition algorithm based on depth image technology was presented which was on the basis of combination of the advantages of Adaboost algorithm and Decision Tree Models on body classification. According to specific application scenarios of personnel evacuation, the sample pool of body shapes was built and the images of positive or negative samples were collected and preprocessed. Adaboost algorithm and CART decision tree were introduced to treat and detect the images of positive or negative samples. The recognition results of three Adaboost classifiers were compared with each other. The results indicated that Gentle Adaboost algorithm was better than the other two. The effects of the parameters of HOG-depth characteristics on the recognition results were analyzed and we found that classification error rate could be decreased by considering positive and negative directions. The effects of traversal means of detection windows on test results were also analyzed. In order to obtain a balance between shorter detection time and more accuracy results, the step should be16pixels. In this case, when the classification time T was equal to3.01s, both the misdetection rate(MR=1.1%) and false detection rate (FR=1.0%) were in ideal conditions.On the last of the paper, a fire supervisory system based on depth image technology was designed and the control method of fire installation based on body recognition was established.