A traffic incident detection method based on depth learning and entropy model relates to the technical field of intelligent traffic. The traffic incident detection method based on depth learning and entropy model includes the following steps: 1) training a convolutional neural network model for traffic incident classification; 2) performing incident classification on input video stream images or video subsegments according to the convolutional neural network model; 3) calculating the entropy value in a period of time according to the result of the incident classification; 4) judging whether atraffic incident occur according to the entropy value. Compared with the prior art, the reliability parameter is calculated according to the proportional and inverse relation between different incident probability, the entropy value and stability. The mutation of the image is analyzed from the global and local characteristics of video images, the occurrence of traffic incidents is detected with the advantages and methods of CNN and the characteristics of mutation detection by entropy model, and the method has characteristics of fast speed and accurate detection.