A crash risk early warning method for work zones based on closed-circuit television system was proposed, in order to improve the timely detection of collision risks within highway work zones. Computer vision techniques were employed for vehicle detection, coordinate transformation, and 3D vehicle shape estimation, enabling the extraction of traffic flow and vehicle information within the work zone. Crash risks in the upstream transition section were predicted based on the traffic features at the starting point of the warning area, utilizing the improved time to accident (ITA) as a quantified indicator for crash risk assessment. By integrating the one-dimensional convolutional neural network (1D CNN), long short-term memory (LSTM) and the attention mechanism, a crash risk prediction model based on the 1D CNN+LSTM+Attention (CLA) framework was constructed. Results showed that the proposed data collection method met the needs for crash risk prediction. The proposed ITA exhibits better sensitivity in risk quantification, compared to other conflict indicators. The CLA-based prediction model demonstrates superior accuracy compared to the LSTM and 1D CNN+LSTM models, with a goodness of fit coefficient and root mean square error of 0.805 and 0.359, respectively. The proposed method can provide crash risk warnings for work zones 90 seconds in advance.