Based on computer vision techniques, the movement tracks of an indicator species (zebrafish) were continuously observed in two dimensions before and after the treatments with a toxic chemical (formaldehyde, 2.5 ppm). Behavioral patterns based on the shape of movement segments were regarded as states, while linear and angular speeds measured from the movement segments were used as observed events for training with a hidden Markov model (HMM). The state sequences were estimated by HMM based on transition and emission probability matrices, and observed events. The movement tracks were further reconstructed based on behavior state sequences generated by HMM. Subsequently, permutation entropy and fractal dimension were calculated to monitor behavioral changes before and after the treatments. Both parameters based on the real and reconstructed data significantly decreased after the treatments, and individual variability was minimized with the parameters obtained from the reconstructed tracks. The parameter extraction based on optimal state sequence by HMM was suitable for resolving the problem of variability in behavioral data, and would be an effective means of monitoring chemical stress in the environment.